<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Dee Carter on Medium]]></title>
        <description><![CDATA[Stories by Dee Carter on Medium]]></description>
        <link>https://medium.com/@dee-carter?source=rss-c1817270117e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*0J5-CqNRdpKrOMJGMCFDuA.jpeg</url>
            <title>Stories by Dee Carter on Medium</title>
            <link>https://medium.com/@dee-carter?source=rss-c1817270117e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 14:20:04 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dee-carter/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[(Paper Review) Motivational Profiling of League of Legends Players]]></title>
            <link>https://medium.com/game-science/paper-review-motivational-profiling-of-league-of-legends-players-7edff8ad22b6?source=rss-c1817270117e------2</link>
            <guid isPermaLink="false">https://medium.com/p/7edff8ad22b6</guid>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[game-development]]></category>
            <dc:creator><![CDATA[Dee Carter]]></dc:creator>
            <pubDate>Tue, 08 Aug 2023 13:11:56 GMT</pubDate>
            <atom:updated>2023-08-08T15:49:52.670Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fbMXjTn1K8kvdAz9Q35Kog.png" /><figcaption>Image source: Midjourney/Author</figcaption></figure><p>This is a review of the journal article “Motivational Profiling of League of Legends Players” by Brühlmann et al., 2020 [1].</p><h3>Overview</h3><p>The authors use an Organismic Integration Theory (OIT) based questionnaire called the User Motivation Inventory (UMI)[2], which is designed to measure player motivational regulations. After gathering 750 participants from the League of Legends subreddit (predominantly 17–25 year old males) they used the responses to this questionnaire to identify motivational profiles using Latent Profile Analysis (LPA)[3]. Responses to several other Self-Determination Theory (SDT), wellness, and experience based questionnaires, as well as in-game behaviour metrics, were also collected. These were aggregated and compared between the four different identified motivational profiles.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FzUFgEU8TsdPvFyvfz4mcw.jpeg" /><figcaption>Figure 1. Originally adapted from Deci and Ryan (2002)[4].</figcaption></figure><p>The authors find statistically significant differences in <em>play experience</em> based measures (how the player perceives the play experience and how it affects them) between the four identified profiles. They also find some statistically significant differences in <em>play behaviour</em> (what the players actually do when playing the game) between the four profiles.</p><h3>Research Problem &amp; Gap</h3><p>Motivation, and SDT in particular, is a key research area within games research <a href="https://www.zotero.org/google-docs/?IiN2YW">[5]</a>, but literature investigating how the motivational regulations of players affects the play experience and in-game behaviours is lacking.</p><p>There is also limited empirical work investigating how differences in motivational regulation affects psychological needs satisfaction. In other words if a player feels amotivated, extrinsically motivated, or intrinsically motivated, how does this impact the players’ psychological needs for competence, autonomy and relatedness when interacting with videogames? This paper aims to address these questions.</p><h3>Contribution(s)</h3><ol><li>The authors utilise a method to cluster responses to the User Motivation Inventory (UMI) questionnaire to identify four motivational regulation profiles of League of Legends players via a process called Latent Profile Analysis (LPA).</li><li>These profiles are correlated with several other questionnaire responses designed to measure interest/enjoyment and pressure/tension, needs satisfaction, wellbeing, positive and negative associations, goal achievement and harmonious and obsessive passion. Doing so the authors show statistically significant differences in experience between players of different motivational profiles.</li><li>The profiles are also used to compare aggregated in-game behaviour metrics between profiles, demonstrating some statistically significant differences in game playing behaviour (albeit less significant relative to the reported experience differences).</li><li>This empirically validates that motivational regulations are related to the play experience and in-game behaviours of players of MOBA games.</li></ol><h3>Methods</h3><p>Several SDT, wellbeing, and other psychological measures (questionnaires) are utilised where response are given using mostly answered using 5 or 7-point likert scales:</p><ul><li>User Motivation Inventory (UMI): a questionnaire based on the Organismic Integration Theory subtheory of SDT. This measures 6 factors including: amotivation, external regulation, introjected regulation, identified regulation, integrated regulation and intrinsic motivation.</li><li>Player Experience Needs Satisfaction (PENS)<a href="https://www.zotero.org/google-docs/?CZ79QE">[6]</a>: a post-hoc SDT based questionnaire measuring needs satisfaction for competence, autonomy and relatedness.</li><li>Vitality<a href="https://www.zotero.org/google-docs/?pM3Tdk">[7]</a>: a measure intended to capture impacts on wellbeing (positive or negative).</li><li>Intrinsic Motivation Inventory (IMI)<a href="https://www.zotero.org/google-docs/?DZDpEA">[8]</a>: the authors specifically adopt the interest-enjoyment and pressure-tension dimensions of the IMI. Authors hypothesise that these are experienced regularly by MOBA players.</li><li>PANAS<a href="https://www.zotero.org/google-docs/?mEiytE">[9]</a>: a measure intended to capture positive and negative affect. Authors hypothesise that these are experienced regularly by MOBA players.</li><li>Achievement Goals<a href="https://www.zotero.org/google-docs/?DH3TK7">[10]</a>: measures a player’s achievement goals orientation (one of: mastery approach, mastery avoidance, performance approach and performance avoidance).</li><li>Harmonious and Obessive Passion (HOP)<a href="https://www.zotero.org/google-docs/?z9lIcd">[11]</a>: a measure to identify autonomous and self-determined internalisation which does not interfere with other areas of life (harmonious), as well as identify non-self-determined internalisation due to external or internal pressure which may interfere with other areas of life (obsessive).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*clcipy8AvtYlJln9qCBVgA.jpeg" /><figcaption>Figure 2 (Table 1 from the paper).</figcaption></figure><p>In-game behaviour metrics are also collected via the League of Legends API using participants’ provided summoner names (usernames). These aggregated metrics included time played, level, total matches, kills, assists, win rate, etc.</p><p>Latent Profile Analysis (LPA) is conducted on the responses to the UMI questions in order to identify four different motivational regulation profiles of the LoL players. Bayesian Information Criterion (BIC), Integrated Complete-data Likelihood (ICL) and Bootstrap Likelihood Ratio Test (BCLR), as well as authors’ own understanding of the profiles’ theoretical conformity were used to identify the appropriate number of profiles from the data.</p><p>The reason for conducting this Latent Profile Analysis was to produce profile labels which enable multi-dimensional analyses. For example, consider if one participant answered the intrinsic motivation (IMO) questions of the UMI very highly, but also answered the amotivation (AMO) questions highly. This participant may be very dissimilar to another participant who responds highly for IMO but low for AMO. The first participant may belong to the Amotivated profile, whereas the second would almost certainly belong to the Intrinsic profile. If we only considered high respondents for IMO as intrinsically motivated players, we would consider these two players the same. Hence clustering methods like LPA are necessary to perform more nuanced, multi-dimensional, and person-centric analyses. People are not one-dimensional, so our analyses should not be either!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qsZGKWs8Vyhkq0B75QNM-A.jpeg" /><figcaption>Figure 3 (Figure 2 from the paper).</figcaption></figure><p>The four profiles (and their distribution of values on each latent variable of the UMI) are shown above in Figure 3. Notice how Amotivated and External are similar, and Intrinsic and Autonomous are also similar. We will discuss these similarities later.</p><h3>Results</h3><h4>Behavioural Game Metrics</h4><p>The authors conducted a correlation analysis between the profiles and aggregated in-game metrics. These tell us how players who fit into each profile behave in the game on average.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bd09b8apFvSdxKnFH4Y4Yw.jpeg" /><figcaption>Figure 4 (Table 3 from the paper).</figcaption></figure><p>Some interesting results pop up. First is that there aren’t many significant differences in the behaviour of the four different profiles. The main significant behavioural differences are: amotivated and autonomous players have higher mean winrate in unranked games, autonomous players have higher mean KDA in ranked games, and autonomous players have higher mean assists in unranked games. The remaining differences are not statistically significant.</p><p>Combined with the similarity of the profiles as per the UMI factors (Figure 2 above), the question arises as to why these profiles seem so similar. We’ll tackle this question shortly.</p><h4>Player Experience</h4><p>The authors also aggregated the results for responses to the other measures (questionnaire questions) by profile to see what the difference in player experience was between each profile. These tell us various things such as the extent to which the players experienced enjoyment, tension, feelings of positive/negative well being, psychological needs satisfaction and so on. The results are more interesting and are summarised below in Figure 5.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mbfIzzLhR0TMbUy-OZCZnQ.jpeg" /><figcaption>Figure 5 (Figure 3 from the paper).</figcaption></figure><p>Describing this plot from left to right. All profiles reported high enjoyment, with the Intrinsic profile reporting highest levels of enjoyment. This is consistent with Organismic Integration Theory (see Figure 1), which posits that more integrated external regulations have more significant positive effects on wellbeing, engagement and enjoyment of an activity. Intrinsic motivation is believed to be the superior form of motivational regulation producing the highest levels of engagement, wellbeing and enjoyment.</p><p>Externally motivated players reported highest levels of tension. This may be due to their motivational regulation focussing on external outcomes, such as feeling pressure to perform well in the game to please their teammates, or to realise some internalised expectation of their own performance.</p><p>Intrinsically motivated players reported the highest levels of needs satisfaction for autonomy, which is also consistent with Organismic Integration Theory. There were not large differences in needs satisfaction for competence among the profiles, but Intrinsic and Autonomous reported slightly higher levels. For relatedness, the Amotivated profile reported significantly lower levels of needs satisfaction. This might indicate that Amotivated players do not feel a sense of connection or community with their fellow players whilst playing the game, and possibly avoid others whilst playing. This is consistent with their behavioural metrics: Amotivated players provide the lowest numbers of assists in both ranked and unranked, the highest number of kills in ranked mode, and the highest number of deaths in ranked and unranked. This suggests a maverick/solo playstyle for Amotivated players.</p><p>In terms of goal achievement, a standout result is that Intrinsically motivated players report the lowest levels of mastery and performance avoidance. This suggests that Intrinsically motivated players are not concerned with losing previously acquired knowledge or skills and are not concerned with performing worse than their peers or some normative standard. These goals are significantly higher for externally regulated players, from which we can infer they play due to an external or internalised pressure to perform highly and not lose their skill.</p><p>This result is supported by the significantly higher responses for obsessive passion for External regulated players. While not a high average response, it suggests their play habits might be more likely to interfere with other aspects of their lives in an unhealthy way. In contrast, intrinsically motivated players report the highest levels of harmonious passion, which is defined as “the autonomous and self-determined internalization of an activity into one’s identity (Vallerand et al., 2003), whereby the activity is aligned with different areas of a person’s life”.</p><p>The responses for Vitality support this finding further, where the Intrinsic profile reports the highest levels of Vitality and External reports somewhat lower levels of vitality. Vitality in this instance is defined as “positive feelings of aliveness and energy … hypothesized to reflect organismic well-being.”<br>However, the Amotivated profile reports the lowest levels of vitality, suggesting that disengagement and lack of understanding of why one plays LoL is somewhat detrimental to well-being.</p><h3>Summary and Critical Reflection</h3><p>In this paper the authors have demonstrated that different types of motivational regulation can relate to the player experience and in-game behaviour. The authors highlighted the distinction between motivational regulations (OIT) and psychological needs satisfaction (SDT), and how these two interact during gameplay. Their empirical analysis is extended to include measures of wellbeing, demonstrating that the different types of motivational regulation also impact players’ experience differently, with intrinsic motivation and more integrated forms of external regulation being the most desirable for player wellbeing.</p><p>The side-by-side comparisons of so many psychological measures in one study of video game players is extremely interesting and applicable to industry.</p><p>However, the sampling method may have biased the results towards players of a specific demographic (17–25 year old males) who are probably all highly engaged (interacting outside of the game on the subreddit is a strong indicator of overall engagement). This might explain why the four identified motivational profiles are very similar, in terms of the underlying UMI measure, additional measures, and in-game behaviours. Casting the net wider to find participants with varied levels of engagement might yield different results with more significant differences between the types of players.</p><p>In addition to the above, this use of gameplay behaviour metrics aggregated over the lifetime of the player assumes that a player’s motivational profile is <em>static</em> over the course of their interactions with a game. This is a big assumption. It assumes that motivational regulations are <em>traits</em> rather than <em>states</em> of the player. In other words: is it possible for a player to change from one motivational type to another over the lifetime of their engagement with a game, or even during the same play session? Organismic Integration Theory asserts that this is the case. If it is true, then comparing lifetime aggregated behavioural metrics between the profiles is less likely to reveal significant differences in behaviour.</p><h3>References</h3><p>[1] F. Brühlmann, P. Baumgartner, G. Wallner, S. Kriglstein, and E. D. Mekler, ‘Motivational Profiling of League of Legends Players’, <em>Front. Psychol.</em>, vol. 11, 2020, Accessed: May 12, 2023. [Online]. Available: https://www.frontiersin.org/articles/10.3389/fpsyg.2020.01307</p><p>[2] F. Brühlmann, B. Vollenwyder, K. Opwis, and E. Mekler, ‘Measuring the “Why” of Interaction: Development and Validation of the User Motivation Inventory (UMI)’, Apr. 2018. doi: 10.1145/3173574.3173680.</p><p>[3] D. Spurk, A. Hirschi, M. Wang, D. Valero, and S. Kauffeld, ‘Latent profile analysis: A review and “how to” guide of its application within vocational behavior research’, <em>J. Vocat. Behav.</em>, vol. 120, p. 103445, Aug. 2020, doi: 10.1016/j.jvb.2020.103445.</p><p>[4] <em>Handbook of self-determination research</em>. in Handbook of self-determination research. Rochester, NY, US: University of Rochester Press, 2002, pp. x, 470.</p><p>[5] A. Tyack and E. D. Mekler, ‘Self-Determination Theory in HCI Games Research: Current Uses and Open Questions’, in <em>Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems</em>, Honolulu HI USA: ACM, Apr. 2020, pp. 1–22. doi: 10.1145/3313831.3376723.</p><p>[6] R. M. Ryan, C. S. Rigby, and A. Przybylski, ‘The Motivational Pull of Video Games: A Self-Determination Theory Approach’, <em>Motiv. Emot.</em>, vol. 30, no. 4, pp. 344–360, Dec. 2006, doi: 10.1007/s11031–006–9051–8.</p><p>[7] R. M. Ryan and C. Frederick, ‘On Energy, Personality, and Health: Subjective Vitality as a Dynamic Reflection of Well-Being’, <em>J. Pers.</em>, vol. 65, no. 3, pp. 529–565, 1997, doi: 10.1111/j.1467–6494.1997.tb00326.x.</p><p>[8] E. McAuley, T. Duncan, and V. V. Tammen, ‘Psychometric Properties of the Intrinsic Motivation Inventory in a Competitive Sport Setting: A Confirmatory Factor Analysis’, <em>Res. Q. Exerc. Sport</em>, vol. 60, no. 1, pp. 48–58, Mar. 1989, doi: 10.1080/02701367.1989.10607413.</p><p>[9] D. Watson, L. A. Clark, and A. Tellegen, ‘Development and validation of brief measures of positive and negative affect: the PANAS scales’, <em>J. Pers. Soc. Psychol.</em>, vol. 54, no. 6, pp. 1063–1070, Jun. 1988, doi: 10.1037//0022–3514.54.6.1063.</p><p>[10] A. J. Elliot and H. A. McGregor, ‘A 2 X 2 achievement goal framework’, <em>J. Pers. Soc. Psychol.</em>, vol. 80, no. 3, pp. 501–519, Mar. 2001, doi: 10.1037/0022–3514.80.3.501.</p><p>[11] R. J. Vallerand <em>et al.</em>, ‘Les passions de l’âme: On obsessive and harmonious passion.’, <em>J. Pers. Soc. Psychol.</em>, vol. 85, no. 4, pp. 756–767, 2003, doi: 10.1037/0022–3514.85.4.756.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7edff8ad22b6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/game-science/paper-review-motivational-profiling-of-league-of-legends-players-7edff8ad22b6">(Paper Review) Motivational Profiling of League of Legends Players</a> was originally published in <a href="https://medium.com/game-science">Game Science</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[There are no “emergent abilities” in LLMs]]></title>
            <link>https://medium.com/better-programming/there-are-no-emergent-abilities-in-llms-2bb42e17ce7e?source=rss-c1817270117e------2</link>
            <guid isPermaLink="false">https://medium.com/p/2bb42e17ce7e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[openai]]></category>
            <dc:creator><![CDATA[Dee Carter]]></dc:creator>
            <pubDate>Sat, 13 May 2023 16:19:24 GMT</pubDate>
            <atom:updated>2023-05-17T15:40:12.572Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0R54hLA199y_GXg6TLqnbw.png" /><figcaption>Image generated by Midjourney</figcaption></figure><p>The 2022 paper titled “Emergent Abilities of Large Language Models” [1] has been the source of much speculation — and much hysteria — about the future of AI [2], [3]. The paper’s central claim is that large language models (LLMs) display “emergent abilities” as they scale to larger and larger parameter counts (hundreds of billions of parameters are common model sizes now).</p><p>In the paper, the authors define emergence as follows:</p><blockquote>“We consider an ability to be emergent if it is not present in smaller models but is present in larger models. Thus, emergent abilities cannot be predicted simply by extrapolating the performance of smaller models.”</blockquote><p>In this article, I’m going to demonstrate why these claims are flawed, and by the end of the article, I’m certain you’ll be convinced that unexpected abilities do not suddenly “emerge” in large language models. The code to reproduce my plots can be found in <a href="https://colab.research.google.com/drive/1bhSjEU0ep6UTGmfaewpapjl163cVNv3L?usp=sharing">this Colab notebook</a>.</p><h3>The Source of the Claim</h3><p>In the paper, the authors use the plots below as the basis for the claims of “unpredictable” and “emergent abilities.” Figure 2 from the paper is the basis of the claims.</p><p>In this article, I’ll scrutinize Figure 1 (shown below) from the paper, which is analogous (and virtually identical to) Figure 2.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/958/1*8JAtSGmX-r3s87oVGfDe9w.png" /><figcaption>Fig. 1 (Source: <a href="https://arxiv.org/abs/2206.07682)">https://arxiv.org/abs/2206.07682)</a></figcaption></figure><p>In these figures, the authors compared results from different model types and sizes on BIG-bench [8] benchmark tasks like arithmetic, word unscrambling, and question &amp; answering.</p><p>As you can see from the plots, there is a sudden and unexpected jump in performance on most tasks as the models scale in the number of parameters. The authors claim that each of these plots shows emergence, and visually, that seems to fit their definition.</p><p>From here on, I will focus on GPT-3 (purple) in Figure 11A since that appears to be the most sudden and steepest increase in arithmetic ability, and so should be a good poster child for the other tasks where emergence is claimed to occur. This arithmetic task (Fig. 11A) shows a particularly extreme-looking example. In this plot, GPT-3 appears to suddenly go from no ability in arithmetic to have some ability at around the 10B parameter mark, and this ability appears to continue to grow rapidly as the parameters scale up.</p><h3>The Problem With Log Scales</h3><p>On to the core issue with the paper: the use of graphs on a semi log scale, and using the visual qualitative properties of these graphs to claim “emergence.” This problem comes from the fact that the x-axis is log scaled.</p><p>For those unfamiliar, a log scale graph uses the log function on one or more axes to scale that axis to better fit the data visually. This is usually done when the data points along one or more axes differ by orders of magnitude, as is the case with these language models: some model data points have one hundred <em>million </em>parameters, and others have over one hundred <em>billion</em> parameters.</p><p>The result of such scaling is that the data points are stretched horizontally to have more even spacing between them, when in reality (on a linear scale), the spacing between the data points would be much much larger. Visually this scaling creates an illusion where a straight diagonal line becomes a curve. Here’s a toy example using simple data points generated at each power of 10 (from 10¹ to 10¹⁴) shown side by side using the same data but on different scales:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/826/1*00eOdhTUaooxAnS3qygb1Q.png" /><figcaption>Fig 2. (Source: Author)</figcaption></figure><p>Now, the point becomes clear. The visualization on the left shows inconsistent spacing between each data point (10¹⁴ is much bigger than 10¹³) but shows that there is a nice linear increase between the input variable x and the response variable y. On the right-hand side, using a log scale on the x-axis, we instead have nice even spacing between each datapoint, but the straight line we observed earlier now appears to be a very dramatic curve! You can imagine this like pulling the leftmost datapoints evenly along to the right.</p><p>This same problem applies to the plots in the “emergent abilities” paper. When viewed with a log-scaled x-axis, the data shows that the performance of LLMs on arithmetic tasks suddenly and drastically increases. However, this is absolutely not the case. Here is the data from the paper plotted in the same way, plotted on a linear scale on the left and a log scale on the right:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/841/1*U-dsLe3zj3B-OaWU9etRjg.png" /><figcaption>Fig. 3 (Source: Author)</figcaption></figure><p>When viewed on a linear scale (left above), the relationship between the number of parameters and the performance of GPT3 on the arithmetic tests seems fairly linear. Indeed the performance increases are not as sudden and unpredictable as the paper claims.</p><p>It should be clear now that the claimed ‘emergence’ is a visual artefact of these log plots and not actually representative of a phenomenon in the underlying data. Given that this plot from Figure 11A in the paper is the example with the most extreme and ‘sudden’ visual increase in performance, performing the same analysis on the other plots would also show that performance increases gradually and not suddenly.</p><h3>Predicting the Unpredictable</h3><p>There’s one remaining issue I wanted to address: the claim of unpredictability. When pressed on their use of log scales, the authors claimed that for this particular test, the <a href="https://openreview.net/forum?id=yzkSU5zdwD&amp;noteId=hxL3GlIjjh">emergence happens after 7 billion parameters</a> for GPT3 [5]. The claim is that the performance of the models before this point is totally random and cannot predict the performance of subsequent model sizes. The authors have also recently followed up with a blog post [6], using this linear plot to argue that the performance between 7B and 13B parameters is unpredictable — and thus demonstrates emergence.</p><p>Let’s test these claims. Using the first six datapoints only — data for GPT3 model sizes up to 6.7B params — we fit a curve to predict the performance of the next largest model: GPT3 with 13B params. Fitting a simple exponential curve to this data and extrapolating yields the following plot:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/562/1*ZgHLop9230EogItCXdEtvw.png" /><figcaption>Fig. 4 (Source: Author)</figcaption></figure><p>The fitted curve is shown as a solid black line. The predicted performance of GPT3 with 13B parameters is shown using the dotted lines. The <em>actual</em> performance of GPT3 with 13B parameters is shown by the orange dot.</p><p>This orange data point wasn’t used to fit the curve; only the first six blue data points were used. This simple model predicts that GPT3 with 13B parameters would have an accuracy on arithmetic tasks of 5.43%, which is not terribly far off the real accuracy of 7.1%.</p><p>What I’m trying to communicate here is that performance clearly doesn’t suddenly and unpredictably increase (i.e., emergence) and is actually quite predictable using simple methods that take only a few minutes to apply. So we have shown that the authors’ claims are wrong.</p><h3>Linear Relationship, After All?</h3><p>Another possibility is that the relationship between model size and performance is actually linear. The variance in performance between model sizes might be due to the stochastic nature of training deep learning models and the large variety of tasks within language training sets. With only one data point per model size, we can’t say for sure, but separate models trained with the same number of parameters would likely have some difference in performance on each task.</p><p>Despite this, fitting a degree-1 polynomial (line) to all available data certainly looks like a linear relationship fits. We could be more confident of this relationship with more of the intermediate data points.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/584/1*NFIFsYRjcUxelNOdkJ8ZUQ.png" /><figcaption>Fig. 5 (Source: Author)</figcaption></figure><p>However, what we certainly cannot do is use the lack of these fine-grained data points at each model size to claim it is impossible to interpolate and accurately extrapolate the performance of LLMs as they scale. In this article, we have shown it is possible and trivial to do so. This remark on the lack of intermediate data points is closely related to the points raised in the recent paper “Are Emergent Abilities of Large Language Models a Mirage?” [7], which I highly recommend for further reading.</p><h3>References</h3><p>[1] J. Wei <em>et al.</em>, ‘Emergent Abilities of Large Language Models’. arXiv, Oct. 26, 2022. doi: 10.48550/arXiv.2206.07682.</p><p>[2] J. Cox, ‘AI anxiety: The workers who fear losing their jobs to artificial intelligence.’ https://www.bbc.com/worklife/article/20230418-ai-anxiety-artificial-intelligence-replace-jobs (accessed May 08, 2023).</p><p>[3] ‘Pause Giant AI Experiments: An Open Letter,’ <em>Future of Life Institute</em>. https://futureoflife.org/open-letter/pause-giant-ai-experiments/ (accessed May 13, 2023).</p><p>[4] ‘BIG-bench/bigbench/benchmark_tasks/modified_arithmetic/results at main · google/BIG-bench,’ <em>GitHub</em>. https://github.com/google/BIG-bench (accessed May 09, 2023).</p><p>[5] J. Wei <em>et al.</em>, ‘Emergent Abilities of Large Language Models’, <em>Transactions on Machine Learning Research</em>, Aug. 2022, Accessed: May 09, 2023. [Online]. Available: https://openreview.net/forum?id=yzkSU5zdwD</p><p>[6] ‘Common arguments regarding emergent abilities’, <em>Jason Wei</em>. https://www.jasonwei.net/blog/common-arguments-regarding-emergent-abilities (accessed May 04, 2023).</p><p>[7] R. Schaeffer, B. Miranda, and S. Koyejo, ‘Are Emergent Abilities of Large Language Models a Mirage?’ arXiv, Apr. 28, 2023. doi: 10.48550/arXiv.2304.15004.</p><p>[8] ‘google/BIG-bench: Beyond the Imitation Game collaborative benchmark for measuring and extrapolating the capabilities of language models.’ https://github.com/google/BIG-bench (accessed May 09, 2023).</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2bb42e17ce7e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/better-programming/there-are-no-emergent-abilities-in-llms-2bb42e17ce7e">There are no “emergent abilities” in LLMs</a> was originally published in <a href="https://betterprogramming.pub/">Better Programming</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Thorough Understanding of Multi-Armed Bandits (Part 2): ε-Greedy and Non-Stationary Problems]]></title>
            <link>https://dee-carter.medium.com/a-thorough-understanding-of-multi-armed-bandits-2-%CE%B5-greedy-and-non-stationary-problems-6127f5f40a91?source=rss-c1817270117e------2</link>
            <guid isPermaLink="false">https://medium.com/p/6127f5f40a91</guid>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[multi-armed-bandit]]></category>
            <dc:creator><![CDATA[Dee Carter]]></dc:creator>
            <pubDate>Tue, 14 Mar 2023 13:41:17 GMT</pubDate>
            <atom:updated>2024-09-12T19:27:59.108Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sgeKQVURlXAe4mTk_TVINw.png" /><figcaption>Source: Author via MidJourney</figcaption></figure><p>Today we’re going to address some of the problems with an ε-first exploration approach for multi-armed bandit problems. In the <a href="https://medium.com/%40dan-carter/a-thorough-understanding-of-multi-armed-bandits-part-1-%CE%B5-first-47f137236ea1">last post</a> we saw how ε-first can perform very well on stationary problems where the true value Q∗​ of each bandit arm (slot machine in our example) never changes. But in the real world we are often faced with problems where the true value of a choice changes over time. In these situations the ε-first exploration approach will not adapt to the changing environment, and will ignorantly keep selecting the same sub-optimal action over and over. As with my previous post, you can follow along and run the code in <a href="https://colab.research.google.com/drive/1cEEQ4UO4K_v8ik-SzKKrVBW25bdMwaaM?usp=sharing">this colab notebook</a>. 📓</p><h3>Non-stationary problems:</h3><p>A non-stationary problem is one where the underlying true value (q∗​) of each bandit arm can gradually change over the course of an episode. Using our slot machines analogy, we can imagine that the slot machines start the day with a random average payout value, and this average payout value can slowly increase or decrease throughout the day. We model this by adapting our <em>BanditProblem</em> class to allow the arm values to change gradually, this is done by simulating ‘random walks’ for each arm, by drawing a small random number from a normal distribution for each arm and adding these to the true value of each arm. So this means the true payouts can gradually go up or down for each arm.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/f204e08542defa0cbbd12cfef2697e6f/href">https://medium.com/media/f204e08542defa0cbbd12cfef2697e6f/href</a></iframe><p>This new logic is handled in the <em>step_arm_values</em> function above, which makes small changes to the true arm values and is called after the bandits are done drawing from the arms.</p><h3>Introducing ε-Greedy:</h3><p>Our first attempt at tackling the non-stationary bandit problem uses the well-known ε-greedy approach. It is fairly simple and similar to ε-first but with a small difference; instead of exploring randomly for some fixed number of steps, ε-greedy explores randomly some % of the time throughout the entire episode. This means ε-greedy never stops exploring some small portion of the time, determined by the value ε. This could be a small number, like 0.01, meaning that the agent explores randomly approximately 1% of the time. It also means that ε-greedy can start exploiting the best found option right away — there is no long period of exploration time needed before exploitation. The incremental average update rule for the q-value estimate stays exactly the same as it was for ε-first, here it is as a reminder:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/230/1*QwOW5xfkyhSOZFzH4GwiCw.png" /><figcaption>Source: Author</figcaption></figure><p>Adding this one simple change — introducing some small chance of exploring randomly every step — is enough to allow the ε-greedy bandit to adapt to non-stationary problems, because it constantly updates its belief about the best choice by some small amount. <br>So is ε-greedy ‘better’ than ε-first? Well, it depends. On a stationary problem where the values of the slot machines never change, then ε-first is probably better (if you can afford the upfront exploration). On a non-stationary problem, ε-greedy will be better.</p><p>Here’s the code for an ε-greedy bandit:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/39be59a8c591f1f183044792cf89eee6/href">https://medium.com/media/39be59a8c591f1f183044792cf89eee6/href</a></iframe><p>Once again most of the logic is handled in the <em>choose_action</em> function. Notice how at each step (choice) we draw a random real number between 0.0 and 1.0, and we exploit the best found arm if that random number is bigger than ε (epsilon). But, if the random number is less than ε, then we explore (choose any random arm). Then we take note of the reward/payout received and update our estimate. There are two update types here, but the weighted one can be ignored for now, I’ll explain it later.</p><h3>Results: ε-First vs ε-Greedy</h3><p>So how does ε-greedy fare against ε-first on a stationary problem? See the graphs below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/744/0*OmonxOKcZJesxqK2.png" /><figcaption>Source: Author</figcaption></figure><p>As expected, on our stationary 10-armed bandit problem the ε-first agent fares better than the ε-greedy agent. This is because the values of the bandit arms never change, so once the ε-first bandit has locked on to the best choice, it exploits that choice continually. In contrast, ε-greedy takes a while longer to find the optimal choice, and even when it does find it, it still explores the other sub-optimal options 10% of the time. The only upside to the ε-greedy approach here is that it starts gathering good rewards almost right away, whereas ε-first takes 1000 exploration steps before collecting high rewards.</p><p>But what about a non-stationary problem — when the values of the bandit arms change? Can you predict what will happen in this case? Which approach will fare better?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/744/0*6mpSbUy_MSNKbmTU.png" /><figcaption>Source: Author</figcaption></figure><p>Aha! Now the tables have turned. While ε-first does for a short moment often find a better choice than ε-greedy right after it is done with 1000 exploration steps, this choice quickly becomes stale as the values of the arms change. ε-greedy fare much better overall, and continues to increase its average score over the course of the episode! However, ε-greedy seems to exhibit the same staleness problem (though not as bad) as ε-first: as the episode goes on it chooses the optimal action less and less. Can you think of a reason for this? Hint: think about the q-value update rule shown above (we discussed this in <a href="https://www.daniel-scott-carter.com/multi-armed-bandits-1-%ce%b5-first/">my last post too</a>).</p><p>The reason is because the update rule puts <em>n</em> in the denominator of the fraction that behaves as the step size, so this step size gets smaller and smaller as the episode goes on. Eventually, this step size will be so small that the q-value estimates will barely change on each update, and their rank (the order of each bandit arm according to our estimates) will almost never change after many episodes. So, eventually, ε-greedy almost becomes exactly like ε-first and gets stuck pulling the same arm when exploiting (but still pulls other arms randomly some % of the time). This explains why ε-greedy slowly makes less and less optimal choices — its arm value estimates are not keeping up with the changes to each arms’ true value q∗ the longer the episode goes on!</p><h3>Recency Weighted Q-Update:</h3><p>To truly solve the stale choice issue for non-stationary bandit problems we need to be a bit more intelligent with our q-value estimate update rule. We saw previously how ε-greedy with a sample average update rule has some problems: the step size gradually gets smaller and smaller. The fix is simple: keep this step size constant! The reason why this works is subtle.</p><p><strong>Theory warning!</strong> The following section will be a bit math heavy. But if you find maths a little dense (I can relate) then I also include a plain English description and accompanying code below. Hopefully that helps! 😄</p><p>So we want to change our update rule to include a fixed step size αα. That’s easy enough — see line (1) below. Line (1) below is all we need to implement this in code, the rest are just there for understanding. The reason why this works — and the key to understanding why it works — as a recency-weighted update (new updates are given more weight) is due to the <em>recursive</em> way αα is applied to the existing estimate Qn​. Realise that Qn​ at any time step is the result of previous updates being applied one after another. Lines (2) to (6) demonstrate this below, where we are essentially unrolling Qn​ as a sum of updates from all previous time steps.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/756/1*I4VpdnRmcRZAL75gQlGwEg.png" /><figcaption>Source: Author</figcaption></figure><p>On line (7) we show this idea in a single line where we express Qn+1​ as the sum of all previous rewards weighted by alpha raised to the power of n−i (the number of steps ago this update occurred). So as we update Q, the contribution of old updates to our current estimate gets exponentially smaller and smaller.</p><p>It’s this attribute of our new recency-weighted average that allows our estimates to stay fresh as the ε-greedy continually explores and the true values of the bandit arms change over time. This update rule reacts much faster to changes in the values of the bandit arms and can do so for as long as the episode continues. Although, if it experiences an unlucky streak of exploration it may be temporarily misled into believing that the optimal action is not actually the best. Even so, it will usually fix this mistake quite quickly when the unlucky streak breaks.</p><p>You can find the code for this update rule in the <em>update_estimate_weighted</em> function in the code snippet above.</p><h3>Results: ε-Fixed vs ε-Greedy vs Recency-Weighted ε-Greedy</h3><p>Naturally, we wouldn’t expect the recency-weighted update ε-greedy to be any better when it comes to stationary problems, and this is true based on our results in the graphs below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/880/0*yWvnTB_OdoAG8-V2.png" /><figcaption>Source: Author</figcaption></figure><p>It’s clear that recency-weighted averages hurt performance in the stationary setting. Whereas ε-greedy eventually approaches choosing the best action 90% of the time (the remaining steps are exploration), the recency weighted ε-greedy chooses the optimal action 80% of the time and does not seems to be improving. In this situation the recency-weighted ε-greedy is limited by the αα value, which determines the ‘forgetfulness’ of the update rule, so this could be tuned a bit to improve performance for the stationary problem setting. But where recency-weighted ε-greedy really shines is in the non-stationary setting:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/874/0*rLMI7SPpNBRKRpKE.png" /><figcaption>Source: Author</figcaption></figure><p>Much better! Adding the recency-weighted update rule allows the ε-greedy agent to outperform both prior approaches on non-stationary problems. The recency-weighted update gradually ‘forgets’ about old updates so that it can quickly switch to the new optimal choice as the episode progresses and the values of the bandit arms gradually change.</p><h3>Discussion and future work:</h3><p>So that’s it, right? We’ve solved the exploration/exploitation problem? Uhm… actually no. In the last two posts we’ve seen some really simple (and quite effective) methods to balance exploration/exploitation, but there’s one more method I want to cover. This method addresses a problem with ε-greedy: when it explores it does so totally at random, but couldn’t there be a way to focus exploration on choices that seem most promising? This is the topic of our next post when I’ll cover Upper Confidence Bound (UCB) action selection, and also some neat tricks to make ε-greedy exploration more effective too! 🤖</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6127f5f40a91" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Thorough Understanding of Multi-Armed Bandits (Part 1): ε-first]]></title>
            <link>https://dee-carter.medium.com/a-thorough-understanding-of-multi-armed-bandits-part-1-%CE%B5-first-47f137236ea1?source=rss-c1817270117e------2</link>
            <guid isPermaLink="false">https://medium.com/p/47f137236ea1</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Dee Carter]]></dc:creator>
            <pubDate>Sun, 26 Feb 2023 09:50:03 GMT</pubDate>
            <atom:updated>2024-09-12T19:28:39.066Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XVLD8f6MY2pTD5KmFDA6ig.png" /><figcaption>Source: Author/Midjourney</figcaption></figure><p>In this post we’re going to discuss the age old problem of making decisions when there is uncertainty. To illustrate what I mean, we’re going to dive right into multi-armed bandits problems and what they are exactly. You can follow along and run the code yourself using <a href="https://colab.research.google.com/drive/1cEEQ4UO4K_v8ik-SzKKrVBW25bdMwaaM?usp=sharing">this google colab notebook</a>. I’ll be updating the notebook to include code and experiments for future MAB posts. ♻️</p><h3>Intro to Bandit Problems:</h3><p>A bandit problem is a situation where we have a series of choices of actions — some of those choices are better than others — but we don’t know which ones. The choices are totally independent of one another, meaning one choice does not affect the next choice, and we can choose from all the options on each turn. For these problems we want to figure out what the best option is as quickly as possible and then continually exploit that option until we run out of turns.</p><p>The classic example is given using a set of slot machines that have different average payouts. We use the term “average payouts” here because there is some randomness (noise) to the payouts, meaning we get a slightly different payout each time. If payouts were certain, we would just need to sample from each machine once to figure out which is best.</p><p>So, ideally we want to figure out which is the best slot machine with the highest average payout as quickly as possible — so we can then focus on playing only on that best slot machine and win as much money as possible. In other words we don’t want to waste turns playing on slot machines where the average payout is lower than the best machine — but first we have to figure out which machine is best. 🎰</p><p>The example we’ll work with from here on is the 10-armed bandit problem — where we have 10 slot machines to choose from, each with different average payouts, and our job is to figure out which is the highest paying machine. We initialise the true value (Q∗) of each slot machine (bandit arm) by sampling from a normal distribution with a mean of 0 and a standard deviation of 1. Code for the problem is supplied below</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/aa79b8aaa4b5ec4fdf0ef8fee9b2f6ad/href">https://medium.com/media/aa79b8aaa4b5ec4fdf0ef8fee9b2f6ad/href</a></iframe><h3>Fixed exploration period (ε-first):</h3><p>Our first and most intuitive solution is to choose some arbitrary fixed period of exploration where we play each machine at random and try to build up some useful estimate of what the average payout is. This strategy is sometimes referred to as fixed-exploration or ε-first (epsilon-first). With this approach we would aim to play on each machine say 100 times, and then average the payouts and use that as our estimate. We would then continually play the machine with the best estimated average payout. Similarly, we could pick randomly 1000 times which machine to play and then base our estimate on the average payouts received. Our estimate is just the sample average from the first ~100 turns on that slot machine.</p><p><strong>Theory warning!</strong> The following section will be a bit math heavy. But if you find maths a little dense (I can relate) then I also include a plain English description and accompanying code below. Hopefully that helps! 😄</p><p>The estimate for each bandit arm (slot) is called a q-value estimate, and the true value for each arm is referred to as q∗​ — this is the thing we’re trying to estimate. We can keep track of this q-value estimate for each arm using a simple update rule derived from the sample average formula:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/357/1*WfbQ6dZWe_8nw2BWKNO7_Q.png" /><figcaption>Source: Author</figcaption></figure><p>where Ri​ is the reward from the current step from a given slot machine and n is the count of the number of times this slot machine has been played. Line (3) follows because (n−1)*(1/n−1)=1, and multiplying the sum by 1 does not change its value. Line (4) follows because:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/387/1*jECKdAQjpgzM1sYQFbx37g.png" /><figcaption>Source: Author</figcaption></figure><p>In plain English, line (6) in the first equation above says that we can update the estimate Qn+1​ by finding the difference between the received reward and the previous estimate (Rn−Qn​) let’s call this δ. We then use this difference to update our estimate Qn​, by moving our old estimate a small step towards the most recent reward. This is done by weighting δ by a step size, defined by dividing it by the number of times this arm has been pulled. We can think of this as taking the form:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/590/1*81qkvUMBx7uwNHf6meJong.png" /><figcaption>Source: Author</figcaption></figure><p>Think of this as a gradually reducing the step size the more times we pull a given arm, because we gradually become more certain of the true value. Notice in equation 1 at line (6), when the denominator n (number of arm pulls) increases, the step size decreases.</p><h3>ε-first code:</h3><p>Okay, enough theory for now, let’s put try putting what we’ve learned into code! 💻</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/3b76c79b39d6883a18d7cbdd7524f61b/href">https://medium.com/media/3b76c79b39d6883a18d7cbdd7524f61b/href</a></iframe><p>You can see line (6) from equation 1 implemented in the above in code in the <em>update_estimate_incremental</em> function. Most of the logic is implemented in the <em>choose_action</em> function, where our agent will randomly explore for some fixed number of <em>exploration_steps</em> and then exploit the best estimated “arm” (slot machine in our example) from then onwards. The FixedGreedyAgent will keep track of the current q-value and number of pulls for each arm, as well as a parameter for number of exploration steps to explore before exploiting.</p><p>All of the code for this blog is available in <a href="https://colab.research.google.com/drive/1cEEQ4UO4K_v8ik-SzKKrVBW25bdMwaaM?usp=sharing">this google colab notebook</a> so you can run it yourself and play around with the parameters. 😃</p><p>I chose to run the experiment for 50000 steps per episode and repeated this 2000 times to get an average score for reach step. Our fixed-exploration greedy agent explores for 1000 steps before exploiting. I recommend using a much lower value for steps and episodes when running on google colab or the experiment will take a long time to complete, but feel free to play around with the exploration steps parameter. We can see the results for the average reward per step for the fixed exploration bandit below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/379/0*TKnp3bntFnt6NhTH.png" /><figcaption>Source: Author</figcaption></figure><h3>Discussion and future work:</h3><p>Okay, so it looks like our fixed exploration strategy worked pretty well, right? Well yes, but there are some problems with this strategy:</p><ol><li>We have to spend X number of turns exploring randomly before we can exploit and get a decent score. This could be very costly in some scenarios, for example if it costs $1 to play each slot machine, and our average return for the first 1000 steps is $0, then it is going to cost us $1000 out of pocket just to explore, before we see any returns! So we may need deep pockets for this exploration strategy.</li><li>Also the more arms we have, the longer we have to spend randomly exploring before we can be confident we have found the optimal one and exploit it. If we had 150 arms or more then 1000 exploration steps may not be adequate. If there are thousands of bandit arms, then this exploration strategy may not be viable at all.</li><li>If we don’t confidently find the optimal choice (slot machine) within those first X exploration steps, then we will lock on to a sub-optimal choice for the rest of the episode! We could be stuck making bad decisions forever.</li><li>If the value of the bandit arms (slots) changes over the course of an episode, our agent will never update its belief about the best arm to choose. It’s fixed decision could get worse and worse as an episode progresses.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/372/0*hsAYLQzeKHICxPnJ.png" /><figcaption>Source: Author</figcaption></figure><p>This plot makes the problems visible. During the first 1000 steps we pick the optimal choice roughly 10% of the time (makes sense; 10 choices picked at random), and after the exploration phase is done the fixed exploration bandit manages to ‘find’ the best option ~95% of the time. But, once the agent is done exploring it never changes its mind about the best option. It also isn’t guaranteed to have found the best slot machine in the first place, especially if the number of arms is considerably more than 10.</p><p>This begs the question: can we explore more effectively? And that is a very, very good question. Actually, that question is so good that people are still trying to figure it out! In the next post we’ll learn about some more simple and effective ways to balance exploration and exploitation, even for non-stationary problems, using an ε-greedy approach (and others), and different q-value estimate update rules. 🤓</p><h3>References:</h3><ol><li>Russel, S. J., &amp; Norvig, P. (2021). <em>Artificial Intelligence: A Modern Approach (Global Edition)</em> (4th ed.). Pearson.</li><li>Sutton, R., &amp; Barto, A. (2018). <em>Reinforcement Learning: An Introduction</em> (2nd ed.). The MIT Press.</li><li>Fragkiadaki, K. (2019). <em>Exploration/Exploitation in Multi-armed Bandits</em>. <a href="https://www.andrew.cmu.edu/course/10-403/slides/S19_lecture6_exploreexploitinbandits.pdf">https://www.andrew.cmu.edu/course/10-403/slides/S19_lecture6_exploreexploitinbandits.pdf</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=47f137236ea1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Incremental Data Science for Mobile Game Development]]></title>
            <link>https://medium.com/data-science/incremental-data-science-for-mobile-game-development-955bb059e8a6?source=rss-c1817270117e------2</link>
            <guid isPermaLink="false">https://medium.com/p/955bb059e8a6</guid>
            <category><![CDATA[game-development]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[games]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <dc:creator><![CDATA[Dee Carter]]></dc:creator>
            <pubDate>Thu, 11 Mar 2021 20:07:16 GMT</pubDate>
            <atom:updated>2021-03-12T21:51:50.936Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wzencV58UCVkpsnzIp53BA.png" /><figcaption>Source: Author</figcaption></figure><h4>Methodically improving the play experience and boosting KPIs</h4><p>Developing games can often feel like searching around in the dark. Sometimes it’s difficult or even impossible to know what you should do next to improve the experience for your players. Or how to nudge those KPIs up a bit. The development team may have great ideas but they often aren’t sure which ones they should work on first, or which ones could have a positive impact. This is where data can come to the rescue.</p><p>In this post I discuss the top 3 ways data science can make an immediate positive impact in mobile game development. These are things all mobile games studios with some analytics capability can do with a live (or soft launched) game. They can also be done completely offline which makes them great for data scientists looking for a first project in a mobile games studio.</p><h3>1. Optimise the FTUE</h3><p>The FTUE is the first time user experience; it’s the crucial first moments[1] when a player first launches the game. All free-to-play games have the same weakness; since they’re free and easy to install, they’re also free and easy to uninstall. I’m sure many people reading this can think of a mobile game they’ve downloaded and then deleted after less than a minute of total play time. Every minute that goes by after installing more and more players will leave[2]. Only a few days after install the majority of your players will have left for good.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/985/0*aS3nvQL2Uzynm97b" /><figcaption>Source: Author</figcaption></figure><p>Any small inconvenience in that time can be enough for a new user to abandon your game, and this effect is most obvious during the FTUE.</p><p>A typical FTUE will give step-by-step instructions to get the player acquainted with the game, resulting in a funnel that looks something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/985/0*jbr8g_OUclp12XoI" /><figcaption>Source: Author</figcaption></figure><p>This is a standard first attempt at analysing the FTUE. We can see a general downwards trend as players leave, but it’s not immediately obvious if there are any <em>major </em>problems. Let’s look at the percentage change from one stage to the next instead.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/984/0*7LIYCyl_c7I84k-s" /><figcaption>Source: Author</figcaption></figure><p>Now the problem areas where users drop out of the FTUE are clearer. The biggest drop comes before finishing the Intro Video, so we could consider changing this. The problem with this conclusion is that the Intro Video takes on average 15 seconds to watch, which is longer than most other FTUE steps. It could just be that more users leave during this stage because there is more time to leave. Let’s find the average time to complete each stage, and calculate a sort-of ‘churn rate’ per second for each stage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/984/0*WsKxFra7nRZVB4he" /><figcaption>Source: Author</figcaption></figure><p>This bar plot tells us a different story, and combined with learnings from the previous plot it gives us some interesting insights. Players don’t just leave during the intro video because there is more time to leave, they also leave at a faster rate than other stages. Perhaps the intro video isn’t exciting enough? Perhaps the audience in this game genre just wants to get straight into the action? The intro video should probably be shortened, or have an option to skip, or be removed outright.</p><p>It also seems like players have less patience during the initial loading screen, so the game programmers could investigate reducing the initial loading time.</p><p>We can simulate what would happen to our KPIs if we managed to reduce churn by half at the Intro Video stage of the FTUE. Maybe by trying one of the previous suggestions. We’re making a bunch of assumptions here, the biggest probably being that churn at the later stages remains unchanged even though churn at the Video Intro stage has improved. In reality this is unlikely to be the case, but just humour me for a while.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/985/0*RnYL_nVr2_Ev8rDD" /><figcaption>Source: Author</figcaption></figure><p>Doesn’t look like much, but we’re projecting a nice uplift in retention even long after the FTUE is over. This is because some of those users who would have quit now stay around for much longer. Day 7 retention for example, is up from 15% to 15.45%, a relative increase of about 2.5%. This translates into a measurable difference in revenue too, assuming an average revenue per daily active user (ARPDAU) of $0.15 that stays constant across the two groups. This is a pretty big assumption that’s also unlikely to be true in reality, we would need to do a live test to know for sure. See below for our 14 day projected change in cumulative revenue for a cohort of 100,000 users:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/985/0*K3bk7QBRAuskbZgJ" /><figcaption>Source: Author</figcaption></figure><p>Now, just think what the impact could be if we fixed the rest of that FTUE… not bad for a day’s work.</p><p>Identifying FTUE issues is possibly the easiest way data can make an impact on early game KPIs. The FTUE is also one of the only places in game development where small optimisations can have large cumulative impacts. In contrast, late-game optimisations usually require much larger changes to make a noticeable difference.</p><h3>2. Modelling Player Churn</h3><p>Retention is widely considered to be the most important metric in free-to-play[3]. As we saw previously, even a slight change in retention can have a large impact on other KPIs and ultimately the success of a game[4].</p><p>Modelling retention (or churn) is a good candidate for a first machine learning project. We can then interpret this model completely offline to identify issues, suggest fixes, and improve the experience for our players[5]. This process can typically lead to some interesting and unexpected suggestions for improvements.</p><p>Tree based models are a good choice since they’re relatively interpretable for us humans, and they typically perform really well. In this example I fit tree based lightGBM models to synthetic game data to predict the probability that a player will churn and never return after the first day (Day 0). Once our model is fit and we are satisfied with its performance we can use <a href="https://shap.readthedocs.io/en/latest/">SHAP</a>[6] to plot the most important features:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/608/0*a8lBjGNz2vJwT196" /><figcaption>Source: Author</figcaption></figure><p>SHAP values use principles from game theory (a branch of mathematics). In a nutshell it measures the contribution of each input feature against each output of a model. In other words SHAP values measure how much a prediction changes with and without each input feature. Above we see the mean absolute contribution of each feature when predicting churn. Higher values don’t necessarily mean higher predicted churn probability, this plot just indicates how influential each feature is on average. This gives us an idea of the most critical areas that we could look into to improve retention in our game. We can also see how high or low values for each feature contribute positively or negatively to our churn predictions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/604/0*YTwfSV-xCQTkIruO" /><figcaption>Source: Author</figcaption></figure><p>Dots are individual users here, with the impact of this feature on the churn prediction along the x axis. The colour indicates higher/lower values for the feature. Some interesting takeaways:</p><ol><li>Players who receive more rewards have a much lower churn probability.</li><li>Players who spend more time playing are much less likely to churn (duh).</li><li>Players with a higher average FPS are less likely to churn.</li></ol><p>Let’s focus on the FPS issue highlighted in point 3. It looks fairly clear that higher FPS reduces the predicted churn probability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/414/0*A8mKezPvbfvVh8vC" /><figcaption>Source: Author</figcaption></figure><p>Using SHAP to plot each individual user, we can get a clearer picture of how the avg_fps value influences predicted churn probability. Here we see that avg_fps values over 50 have a significant impact on reducing predicted churn. We can colour this plot using another input feature to figure out if there are other relationships between input variables at work:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/471/0*94K3G7sCj8WjKeud" /><figcaption>Source: Author</figcaption></figure><p>An interesting finding here is if a player has a low average fps and they also have a high number of deaths per game they are predicted to be more likely to churn. Also interesting is that this relationship starts to reverse for users with over 30fps average. They become less likely to churn if they have higher deaths per game. It turns out that ⅓ of our players have an average fps less than 30. Using these insights we could suggest that the dev team try to reduce the difficulty of the game for devices with low performance, and also see if there are some easy ways to improve the performance of the game in general. Let’s assume that an FPS boost of 10% is achievable for all players (except those already at the 60 FPS cap). We can then estimate what the impact of making this improvement would be by using our model: simply multiply the avg_fps column of our test data by 1.1 and feed this to the model for predictions. Then we can compare the average churn predictions for these two sets, and plots these as retention/revenue curves:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/985/0*zzSZEh1Wc3QlurOM" /><figcaption>Source: Author</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/985/0*7G_E2gxk43vWD8th" /><figcaption>Source: Author</figcaption></figure><p>So we’re projecting a relative improvement of 3.6% in day 7 retention (15.5% to 16%), and a 3.3% increase in revenue after 14 days on top of our previous FTUE improvement.</p><p>This is a bit naive, since we’re assuming that average fps is completely independent of other features in the data when it most likely isn’t. However, this does give us a good estimate for the kind of improvement we would hope to measure in a real A/B test. With this approach we can model virtually any aspect of our game, identify areas with the biggest potential for improvement, and generate testable hypotheses for how significant the impact of making such a change would be on various KPIs. Stakeholders often find these data-driven suggestions quite compelling.</p><h3>3. LTV models for UA optimisation</h3><p>Aside from improvements to the game itself there are also great opportunities for data science to help with user acquisition and performance marketing. UA (user acquisition) usually involves specifying some CPI (cost per install) you are willing to pay when showing ads in a specific country on a specific platform. There are many other ‘buying’ methods depending on the ad network but for didactic purposes I’ll stick with regular CPI here since this is usually what all the others boil down to anyway. The main contribution is giving UA managers early signals and insights so they’re better informed about where spend on advertising is profitable and where it is not, but also to help identify sources with untapped potential. This is really crucial for a mobile game for several reasons:</p><ul><li>The app stores are flooded with games, without good marketing most players will never know your game exists[7].</li><li>Spending on an ad campaign that will never turn a profit will only incur losses.</li><li>Under spending on an ad campaign with high ROI could mean sub-optimal growth.</li></ul><p>Marketing managers will typically use historical data to plot a simple LTV curve[8] and use this to estimate return on ad spend by some given date. This is where data science comes in. Using data from the early play sessions of our players, we can build a more sophisticated and robust model that predicts the long-term value of users. We can then apply that model to new users acquired through advertising to predict their long-term value, and understand much earlier on if each ad campaign is likely to make a profit. These models are generally more robust to changes in the product and fluctuations in install ‘quality’ when compared to simple historical LTV curves.</p><p>Our UA manager has informed us that we are aiming for a 100% ROAS (return on ad spend) by D90 on a particular campaign. Meaning she expects to have achieved break-even for each daily install cohort by D90. A few days after we start the campaign we can use our trained model to predict D90 value for each user acquired by the campaign so far. We can then aggregate this to compare predicted D90 ARPI (average revenue per install) against the campaign cost per install (this is sometimes called an LTV/CAC ratio[9]).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/984/0*dAHnaiNELgzK49yb" /><figcaption>Source: Author</figcaption></figure><p>This is great - our predicted D90 ARPI is higher than our acquisition costs in almost all cases. But not so fast, a lot of information can get lost in averages, and we know that user spend generally follows a pareto-like distribution. Meaning a very small minority of players spend most of the money. Some of these ‘outlier’ players might be inflating our numbers, so let’s skim the 99th percentile of spenders off our prediction data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/984/0*euIwgJppEO-i23LM" /><figcaption>Source: Author</figcaption></figure><p>When removing the top 1% of spenders from our predictions our campaigns look less profitable. What this tells us is that we are reliant on a very small % of installs to make some of our campaigns break even by D90, and we are exposed to high risk of losses due to the day-to-day variance in the number of these high-value users we can acquire through these campaigns. This quickly turns into a discussion about risk tolerance, which is very specific to each studio so I won’t go into it here.</p><p>We can therefore recommend that our UA manager cancel the campaign in France, and reduce the CPI in Britain since that country appears to be the most reliant on ‘whales’ to only just break even by D90. Countries where we could increase CPIs to attempt to achieve more growth are Australia, Canada and Switzerland.</p><p>An important takeaway here is that campaign performance isn’t fixed. It can change for a variety of reasons. With this in mind we should evaluate the performance of our ad campaigns often and project D90 ROAS using our model, and feed back to UA managers to make adjustments. By doing so we can directly contribute to additional revenue made by our game, which we could then invest back into increased development or additional ad spend for more growth. This is the basics of how we use data to turn our game into a growth engine :)</p><h4>References:</h4><p>[1]<a href="https://www.interaction-design.org/literature/article/first-time-use-experience-ftue-start-as-you-mean-to-go-on">https://www.interaction-design.org/literature/article/first-time-use-experience-ftue-start-as-you-mean-to-go-on</a></p><p>[2]<a href="https://medium.com/googleplaydev/why-the-first-ten-minutes-is-crucial-if-you-want-to-keep-players-coming-back-to-your-mobile-game-4a89031b6308">https://medium.com/googleplaydev/why-the-first-ten-minutes-is-crucial-if-you-want-to-keep-players-coming-back-to-your-mobile-game-4a89031b6308</a></p><p>[3]<a href="https://supercell.com/en/news/10-learnings-10-years/7436/">https://supercell.com/en/news/10-learnings-10-years/7436/</a></p><p>[4]<a href="https://www.gamesindustry.biz/articles/2016-09-08-usd10m-production-budgets-avoiding-the-shark-fin">https://www.gamesindustry.biz/articles/2016-09-08-usd10m-production-budgets-avoiding-the-shark-fin</a></p><p>[5]<a href="https://medium.com/googleplaydev/churn-acting-before-its-too-late-a19f6a680c09">https://medium.com/googleplaydev/churn-acting-before-its-too-late-a19f6a680c09</a></p><p>[6]<a href="https://shap.readthedocs.io/en/latest/">https://shap.readthedocs.io/en/latest/</a></p><p>[7]<a href="https://mobiledevmemo.com/why-paid-ua-is-the-primary-driver-of-discoverability-for-mobile-apps/">https://mobiledevmemo.com/why-paid-ua-is-the-primary-driver-of-discoverability-for-mobile-apps/</a></p><p>[8]<a href="https://quantmar.com/17/What-is-lifetime-customer-value-ltv?show=18#a18">https://quantmar.com/17/What-is-lifetime-customer-value-ltv?show=18#a18</a></p><p>[9]<a href="https://mobiledevmemo.com/ltv-cpi-vs-cashflow/">https://mobiledevmemo.com/ltv-cpi-vs-cashflow/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=955bb059e8a6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/incremental-data-science-for-mobile-game-development-955bb059e8a6">Incremental Data Science for Mobile Game Development</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deconstructing Fall Guys — How to Supercharge Organic Installs]]></title>
            <link>https://medium.com/game-science/deconstructing-fall-guys-how-to-supercharge-organic-installs-19a9e3c9ef6?source=rss-c1817270117e------2</link>
            <guid isPermaLink="false">https://medium.com/p/19a9e3c9ef6</guid>
            <category><![CDATA[mobile-game-development]]></category>
            <category><![CDATA[games-design]]></category>
            <category><![CDATA[fall-guy]]></category>
            <category><![CDATA[ux-design]]></category>
            <category><![CDATA[marketing]]></category>
            <dc:creator><![CDATA[Dee Carter]]></dc:creator>
            <pubDate>Sun, 04 Oct 2020 16:14:26 GMT</pubDate>
            <atom:updated>2024-09-12T19:31:22.990Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dCRjypHrc9RKRBYk" /><figcaption>Source: <a href="https://www.mediatonicgames.com/game/fall-guys">https://www.mediatonicgames.com/game/fall-guys</a></figcaption></figure><p>Fall Guys is undoubtedly one of the great gaming success stories of 2020. With over 7 million copies sold on Steam alone, it took the <a href="https://twitchtracker.com/games">top spot on Twitch</a> on several occasions with a peak of 620,158 viewers and became the <a href="https://www.cnet.com/news/fall-guys-becomes-most-downloaded-ps-plus-game-of-all-time/">most downloaded PS Plus game of all time</a>. It’s safe to say that Mediatonic tapped into the illusive casual formula: broad appeal gameplay that is easy to pick up but difficult to master. But with the inevitable decline in interest kicking in and with Season 2 right around the corner, let’s take a look at how Fall Guys became such a huge success, and how it can maintain relevance in a highly competitive space.</p><h3>Design &amp; Mechanics:</h3><p>In Fall Guys the objective is extraordinarily simple: race to the finish through a Mario-like obstacle course against 59 other jelly-bean shaped players. The controls aren’t complex: the usual movement controls, as well as jump, dive and grab. The simplicity of these inputs and the straightforward objectives combine to make a really accessible game where a first time player, or even a first time viewer, can quickly understand what is going on. The depth comes from the emergent gameplay generated by the Battle Royale style multiplayer and the heavy use of in-game physics. Even other players become obstacles to <a href="https://youtu.be/LL70vTa0E2g">Fall ’N’ Roll</a> over before reaching the finish line. It’s a truly chaotic, comic, free-for-all.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2Je-_Wngb2sQnERY" /></figure><p>This randomness and subtle depth of gameplay create memorable and unpredictable experiences that delight and frustrate players, the most famous case being <a href="https://www.gamespot.com/articles/timthetatman-finally-wins-in-fall-guys/1100-6481162/">TimTheTatMan’s tragedy filled quest for the coveted crown</a>. Surprising and unexpected outcomes are among the most effective methods for generating highly engaging experiences, it’s something Battle Royale games, and Gacha systems, are known to exploit to great effect. Fall Guys doubles down on this with just the right amount of chaos baked into the core experience: twirling propellers, balancing bars, see-saws, to name but a few relatively unpredictable game elements. Then there’s the other players: the icing on the emergent cake. Everybody is in with a shot at the crown in every game, and despite the limited set of courses to play through, each game feels challenging and different.</p><p>Fall Guys doesn’t take itself very seriously at all, and that’s infectious — it’s far more accessible because losing is an entertaining part of the experience. Having victory snatched away by a giant swinging hammer doesn’t feel bad because of the sheer ridiculousness of it. These are the moments that players share with their friends and create entertaining content for influencers. Playing with friends only enhances the fun: the rewards for doing so are implicit, rather than explicit, in the form of hilarious and memorable moments. Implicit motivations are the key ingredients when engaging players for the long term, and so Fall Guys has teed itself up for long term success — if it can fully capitalise on social play.</p><p>The visuals follow a bright and over-saturated aesthetic that would make even Haribo flinch. The colours are eye-catching, enough to get your attention in any YouTube thumbnail, and the character designs are cute and endearing. It’s as if the Artists sat down and decided to pump all of the assets through a <a href="https://deepfriedmemes.com/">meme deep fryer</a> to crank up the contrast and saturation to the max before shipping. There’s no doubt this adds to the virality factor of the game; it’s just so pleasing to look at and the ‘game show’ theme makes it enjoyable to watch.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*PnqveHZtEPPdYRej" /></figure><p>Mediatonic deserves Kudos for their UI/UX design: for a PC game this really wouldn’t seem out of place on mobile at all, and actually puts most mobile UI/UX designers to shame. Navigation is simple: the game will often take you from one menu to the next and surface the relevant information cleanly, such as with the rewards flow at the end of each game. Kudos is pretty much the sole currency which boils other rewards systems down into a simpler form.</p><p>My only complaint here is that the main menu needs more work: the top menu navigation bar could use some life, maybe add a call to action on the store button to draw more attention to the store offers and to let the player know that it’s actually a navigable menu. The player character, bottom-left nameplate, and season progress UI elements are not interactive when they really should be. These would make great alternative ways to navigate (that would also happen to be touch screen friendly *wink wink*). Allowing for clicking the player character to go to the character wardrobe, and clicking the season progress bar to navigate to the season menu would be much more intuitive. These UI interactions lend themselves to the affordances already baked into the UI (drop shadows make things look ‘clicky’). Even so, the ‘3-click rule’ for navigation is rarely broken.</p><p>Aside from this the UX of the core game is superb, with clear signals and affordances communicating the purpose of in-game elements, such as arrows on spinning platforms, markers hovering over friends’ heads, doors that look breakable, and checkpoint flags with clear sounds and effects. Ultimately these little details and finishing touches reduce the need for meticulous tutorials, reducing unnecessary confusion and cognitive load on the player, which makes for an extremely intuitive experience right from the first play session. For a much more in-depth and expert UX breakdown I’d really recommend <a href="https://medium.com/@nida.ahmad/the-playful-ux-of-fall-guys-d2455c25f7db">this blog from Nida Ahmad of Netspeak Games</a>.</p><h3>Core Loop:</h3><p>The core loop is straightforward: jump into game show themed matches where each round more and more players are eliminated while attempting to reach the finish line. The higher your placement, the better the rewards (the currency of which is ‘Kudos’, and the equivalent of traditional Exp is called ‘Fame’). Once only a handful of players remain, the final round is initiated, which usually constitutes a slightly more difficult course. The winner of which earns a ‘Crown’ — the second in-game currency which is only obtainable by winning (or as a season progress reward). Once the player is eliminated or the game is over they go through the rewards flow and are dropped back at the main menu, ready to hit play again, or open the store to spend their hard-earned Kudos.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*bligrAQnmWqauhP3" /></figure><p>The core loop follows a fairly standard recipe: play -&gt; reward -&gt; upgrade -&gt; repeat. It’s perfect for shorter play sessions and provides quick turnarounds for player gratification. This short gratification offsets the long droughts waiting for a win/crown. For the most part the player is hand-held through the core loop, guiding the player from one phase to the next, which almost completely removes points of friction and player churn (with the exception of the main menu/store). Excluding the gameplay itself, there are only two interactions in the core loop: a play button to join a game and a button to confirm rewards when the game is finished that takes you right back to the main menu, ready to go again. There is one additional interaction to equip/confirm a new reward if the player levels up during the end of game rewards flow, but this occurs infrequently and ensures the player is aware of their free reward. It also reminds the player regularly that they’re making progress. <strong>Progress is a strong motivator</strong>.</p><p>The main points of friction occur in the menus: store, cosmetics wardrobe and season progress. Despite it being relatively easy to navigate to these menus from the main menu/lobby, this process could still be improved and simplified as mentioned previously. Alternative ways to navigate menus can only reduce friction for players whose intuition differs from what the UI Designer expects. A call to action on the store button would also draw more attention to the store deals and the fact that the top bar is a navigation bar.</p><h3>Progression &amp; Monetisation:</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*auCMo-ZvQvJB-0_R" /></figure><p>Fall Guys is a mix of premium and free-to-play in terms of monetisation. For Steam users the game costs $19.99, but for PlayStation Plus users the game was free for the first month after release. It’s certainly not a AAA game, hence the lower price tag, but what Mediatonic seem to have grokked (probably due to their experience with mobile free-to-play) is that there are open pastures in the Console/PC space for casual and core game experiences with a <em>very </em>light sprinkle of free-to-play style monetisation. The Console/PC audience is one that has traditionally been hostile towards ‘micro-transactions’, but games like Fortnite have since proven that there is a huge addressable market on HD platforms for games with a low (or non-existent) initial cost. This generally means avoiding anything that might be perceived as pay-to-win or gambling, and not gating content that ‘should be available at launch’ behind additional payments.</p><p>Even with these guidelines in mind, Mediatonic has taken the most player-friendly approach possible: offering cosmetics in exchange for Kudos which can be purchased or earned simply by playing the game. An average player could easily earn enough Kudos to afford any of the shop cosmetics after a few games. Kudos can also be purchased, and the prices are very fair compared to other games: $4.99 bags you 12500 Kudos, which can buy several cosmetics. What this means in practise is that each cosmetic item can be optionally purchased with real money for around one dollar. It’s a notable departure from more aggressive monetisation on mobile, where one or two gacha pulls might cost $4.99. Crowns are the other currency used for buying cosmetics, which cannot be purchased with real money. Crowns can only be earned by winning Shows, so plenty of the most interesting cosmetics are reserved as rewards for playing and doing well. A limited number of exclusive cosmetics are also only available in the DLC store as separate purchases.</p><p>In-game monetisation of premium games, especially those on PC and Consoles, is therefore (generally) an act of balancing fair, non-pay-to-win, and high value-prop purchases with the need to generate a reasonable amount of revenue. Tricky business. Fall Guys clearly errs far on the side of caution but there is room for additional monetisation without damaging player sentiment. One idea would be to add a battle-pass style premium track to the existing Season progress, you can check out <a href="https://medium.com/@dan_carter/the-battle-pass-how-to-get-it-right-74b0a4041eff">my previous article</a> on best practises when implementing battle pass monetisation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gghbPcXimxJ3mdIS" /></figure><p>The main progression system is implemented as a free battle pass. Every game the player earns Fame (read: Exp) that carries them towards the next Season level. At each level there is another free gift to reward the player for their efforts. With 40 levels total (39 with rewards) the cumulative value gifted to the player is approximately $38.40, plus 11 Crowns which don’t have an approximate ‘real-world money’ value because they can’t be purchased. This leaves lots of headroom for a premium option which offers higher value rewards for a cost that’s in line with the other available purchases.</p><h3>Retention &amp; Engagement:</h3><p>Long term engagement and retention are probably the biggest areas that Fall Guys could work on. I have no doubt their early engagement/retention numbers have been stellar, in no small part thanks to their excellent UX, but to guarantee long term success Fall Guys is going to need some work. The season progress system just isn’t enough, and unfortunately Fall Guys doesn’t have the depth or style of gameplay that might encourage long-term mastery and competitive play.</p><p>As hinted at before, I think the first thing Mediatonic could do to boost retention would be to really double down on social play. The fastest way I could imagine them doing this would be to add an Exp/Fame boost when playing with friends, to reward socialites with faster progress. Second, I would suggest setting up a separate playlist for team modes only (Fall Ball, Egg Scramble and so on), these modes are seriously the most fun when playing with a group, and I’m convinced this is the area where Fall Guys shines the most. It’s a bit sad when you are the only one out of your friends to not qualify; sure you can sit and watch to cheer on your friends, but it feels very much like you’re playing separately. Being left behind sucks, and leaving a friend behind while you play the next rounds also sucks. You get no rewards if you quit a round because your buddy didn’t qualify, so it makes more sense to go out together (this is probably the reason why you’ll often see players waiting around before crossing the finish line, or throwing themselves off a ledge when a new round starts). Team modes wouldn’t have that problem. The game is so fun when playing with friends, and a teams only playlist really wouldn’t take much effort (amateur UI suggestions below).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZIz6LRezMu2-FXBN" /></figure><p>Another suggestion to help short and mid-term retention would be random daily and weekly challenges. Think things like: “qualify in Block Party”, “grab 5 players”, “steal 5 tails”, “smash 3 fake doors”, etc. Challenges are a great way to encourage player engagement since they provide not only more things to do, but more routes for progression. It encourages players to come back daily to guarantee some extra Exp to help them along. For that reason they are often paired with a much stronger long-term retention driver, such as a battle pass, base building, leaderboards, or clan challenges. Without such a pairing with a grander objective to give them purpose, daily challenges often fall flat. As the number of minigames increases they can also be added to a daily/weekly rotation to help keep the game feeling fresh from day to day, and provide more opportunities for time-limited daily challenges.</p><p>For longer term engagement (up to one month), the Season pass can be developed further to provide some real incentives for players to complete it. A really cool cosmetic set for reaching the end of the season progress/battle pass is a good place to start. As of right now the final season cosmetic is a victory pose that can’t be viewed to see what it actually does, and many players won’t ever get the opportunity to use it anyway since it requires <em>winning </em>to see it in action. The victory poses seem a little bit too exclusive, so they aren’t a good incentive. Of course more interesting cosmetics, or even toys (throwable tomatoes anyone?) could also be introduced.</p><p>For really long-term retention Fall Guys needs to do a bit of soul searching. The party game theme doesn’t lend itself to key player motivations: Competence, Autonomy and Relatedness. The game isn’t suited to competitive play beyond initial mastery (Competence), nor does the gameplay provide a multitude of creative possibilities or choices for players (Autonomy), nor does it maximise the potential for meaningful social interactions to satisfy the player need to feel like part of a group (Relatedness). It would be challenging for Mediatonic to tackle any of these areas, however, if they really have the budget and ambition, and want to commit to Fall Guys long-term, then there are plenty of ideas worth exploring: a level creator &amp; community creations playlists, casual in-game competitions and prizes, doubling down on communities and promoting them in-game, etc.</p><p>Finally, LiveOps has plenty of potential in Fall Guys as players’ FOMO (fear of missing out) is fantastic for increasing interest and engagement. It’s an approach well suited to this type of game, and I’m sure Mediatonic could do a great job of conjuring excitingly-themed live events to entertain and amuse their fans.</p><h3>Community &amp; Marketing Strategy:</h3><p>If there’s one area where Fall Guys really shines, this is it. Amassing such a huge following and generating marketing buzz is no easy feat, but Fall Guys really nailed it. This is the area where Mobile can learn the most, since the majority of top grossing mobile games are reliant upon performance marketing for scale. But in order to make a game launch as successful as possible, and make subsequent growth more robust, we can learn from the best and develop organic marketing strategies.</p><p>Oliver Hindle (the Fall Guys community manager) put together <a href="https://threadreaderapp.com/thread/1297200779565170690.html">an incredible twitter thread on the Fall Guys community/growth strategy</a> that I’d <em>highly </em>recommend you check out (<a href="https://www.gamesindustry.biz/articles/2020-09-25-the-social-media-strategy-behind-fall-guys-ultimate-knockout">and later an article on gamesindustry.biz</a>). There are things in there that every developer can learn from. One thing the team at Mediatonic did incredibly well was to identify how their game could go viral by understanding how it would be perceived by its <em>real </em>audience. They then tailored their strategy towards that profile to generate maximum buzz. To do so, they needed outreach (Twitter) to bring in new community members, and also places for these community members to mingle and discover one another (e.g. Discord. Reddit is another excellent example).</p><p>Part of the Twitter strategy for drawing in new community members was to adopt the culture and language of these users. The Twitter account was clearly given a lot of autonomy to do this, and to players it stood out as ‘one of their own’ from the crowd of robotic corporate accounts. Once engaged they focused on enabling these community members as much as possible, by promoting user generated content, memes, and in-jokes, which nurtured an engaged community and effectively outsourced the marketing to them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ocvQPGXXq4bMt1st" /><figcaption>Source: Twitter, Oliver Hindle, @OliverAge24</figcaption></figure><p>Pairing the community strategy with a savvy influencer strategy paid dividends. Mediatonic reached out to over 1000 influencers before launch, and enabled their new community marketing army to bring in additional influencers organically. It worked: Fall Guys stormed to the top of Twitch with over 600,000 concurrent viewers shortly after launch, and it kept gaining momentum. Making this process as smooth as possible is key: set up a form so influencers can sign up and have ‘Press Kits’ ready to go that anybody can access. Nobody should have to ask you for anything in order to stream/write/make a video about your game, it should all be there. Rather than asking influencers to stream/play, another effective way to get them interested in your game is to involve them in the development somehow. <a href="https://venturebeat.com/2017/03/30/space-ape-games-taps-20-youtube-influencers-for-fastlane-road-to-revenge/">SpaceApe’s “Fastlane: Road to Revenge” is a great example of this approach</a>. Frogmind’s “Rumble Hockey” is another more recent example which <a href="https://www.pocketgamer.biz/news/73224/rumble-hockey-revenue-share/">offers a 100% revenue share to content creators</a>. Clearly there are endless ways to be inventive with an influencer strategy.</p><p>After launch Mediatonic’s social media pounced on every opportunity they got to engage with influencers and creators. TimTheTatMan is a great example of a popular streamer who the community team targeted with memes and jokes that went viral — the mischief caused headlines and drew in even more players.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ABfD-jwDHb8zGzdh" /><figcaption>Source: Twitter, Oliver Hindle, @OliverAge24</figcaption></figure><p>Of course, Fall Guys was also offered for free to PS Plus subscribers for the launch month, which drove massive interest from the beginning. Having big promotions, or ‘featuring’, is a sure fire way to boost organics for a short time, and it’s an obvious area of growth for Fall Guys. Doing big deals to get the game onto different platforms (Xbox Game Pass? Switch? App Stores?) and reaching as many players as possible makes a lot of sense. If they do go down the Mobile route then polishing the store page with good App Store Optimisation (ASO) practises is another way to maximise organic installs and hit trending pages.</p><p>All in all, Mediatonic have done a terrific job with Fall Guys but it remains to be seen if they can fully capitalise on the enormous player base they’ve drawn in. The game has started to see a decline in interest, but with some smart investment in developing the right areas, releasing the game on more platforms (perhaps some live events to coincide with this), and continuing to mobilise their huge social media following, Fall Guys could certainly extend its success well into the future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=19a9e3c9ef6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/game-science/deconstructing-fall-guys-how-to-supercharge-organic-installs-19a9e3c9ef6">Deconstructing Fall Guys — How to Supercharge Organic Installs</a> was originally published in <a href="https://medium.com/game-science">Game Science</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Battle Pass — How to get it right]]></title>
            <link>https://medium.com/game-science/the-battle-pass-how-to-get-it-right-74b0a4041eff?source=rss-c1817270117e------2</link>
            <guid isPermaLink="false">https://medium.com/p/74b0a4041eff</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[battle-pass]]></category>
            <category><![CDATA[analytics]]></category>
            <category><![CDATA[games]]></category>
            <category><![CDATA[game-development]]></category>
            <dc:creator><![CDATA[Dee Carter]]></dc:creator>
            <pubDate>Mon, 01 Jun 2020 17:19:34 GMT</pubDate>
            <atom:updated>2020-06-01T18:11:34.918Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Battle Pass — How to get it right</h3><p>The battle pass — these days it seems like everyone wants one. DOTA 2 were the first to try it in 2013, but it really exploded onto the scene with the Fortnite battle pass: the ‘skins’ and V-bucks sent tweens into a frenzy all over the western hemisphere. The massive success of Fortnite and the battle pass inevitably led to interest from the rest of the free-to-play world, mobile in particular, and it wasn’t long until battle passes appeared in Apex Legends, Gardenscapes, Call of Duty, Archero, Rocket League and more. It reminds me of a scene in Toy Story (if you know me personally you’ll know I quote this movie all the time):</p><blockquote>“If you don’t have one, GET ONE.” — Woody et al, 1995.</blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FbRCV0uer1Uc%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DbRCV0uer1Uc&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FbRCV0uer1Uc%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/d3a293754e4c6cd7826a05b017602983/href">https://medium.com/media/d3a293754e4c6cd7826a05b017602983/href</a></iframe><p>You may be thinking this could be just the thing <em>your </em>game needs, but hold on there champ, let’s not get ahead of ourselves. It’s easy to be drawn in by the allure of potentially higher revenues, but the battle pass is shockingly misunderstood and the nuances that make for successful implementations are almost ubiquitously overlooked. So let’s slow down for a moment and take a deep dive into the battle pass recipe for success, what you should be doing to implement it, and when you should avoid it.</p><h3>Fortnite vs Archero:</h3><p>Take a look at the Fortnite implementation for example; the catalyst for industry wide adoption. It not only works as an excellent monetisation/conversion mechanic, but a fantastic and clever retention/engagement mechanic too. Let’s drill down into this particular implementation — because it’s probably the most well known, and easily the best implementation of a battle pass out there. As we go, let’s compare the Fortnite battle pass to the one found in Archero, and demonstrate why the former is clearly better than the latter.</p><h4>Boosting Conversions:</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*IUtc7JJ_14nyKVtG" /></figure><p>One thing that is generally well understood about the battle pass is that it can boost conversions significantly. As a player progresses up the free tiers of the battle pass, they earn free rewards and also unlock premium tier rewards that are only accessed by buying the battle pass. It creates a progressively better and better value proposition to the user the more that they play; the player has <em>already </em>earned the rewards, they just need to pay to unlock them. It’s a similar idea to the ‘Piggy bank’ monetisation mechanic; the more you play, the more attractive the IAP becomes. Most product/designer folk are well aware of this, but I really want to explore the nuances that many of them miss.</p><p>At level one, the paid tier of the Fortnite battle pass offers the most when compared to any other battle pass level at approximately $20 to $35 value. The real-money cost of the battle pass is $9.50. This means that if the player buys the battle pass at level 1 they are still rewarded with an immediate return between 2x and 4x value. This first battle pass tier usually includes two unique skins (purely cosmetic avatars) which <em>cannot be obtained via any other means</em>. These skins will never be available again after this battle pass expires: they cannot be purchased or earned in any other way or at any other time. This is an important point:</p><blockquote>the most significant perks you’re offering in your battle pass must be unique and available only for the duration of this pass, in order to maximise perceived value and exclusivity.</blockquote><p>A player who ‘completes’ the Fortnite battle pass before the season ends will receive a 48x value return on their $9.50 purchase, netting the player approximately $450 worth of swag (is swag still a word kids use?). This is a staggering value offer that dwarfs pretty much any other battle pass out there.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*fSiEIa3w99uhogU9aGQStw.png" /><figcaption><a href="https://dscarter95.github.io/medium_data_viz/archero_fn_cumulative_value.html">https://dscarter95.github.io/medium_data_viz/archero_fn_cumulative_value.html</a></figcaption></figure><p>Compare this to Archero, whose Pass costs $4.99 and the value of the first pass level is ~$7.20, or approximately 1.4x value for the first level. It’s just not as enticing right from the get-go. Instead of rewarding a purchaser with value up front, Archero opts to reward them with an Epic Item for reaching the end of the pass. Without this final boost the Archero pass would be worth significantly less overall (roughly 30% less). The total value of the Archero pass is roughly $150, and so it provides much less value (30x initial purchase) compared to the Fortnite pass (48x initial purchase) when completed. This is largely due to the Archero pass being much shorter in duration (23 levels, new season every 14 days) compared to Fortnite’s pass (100 levels, 10 weeks approx.).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/317/1*yOQJldaujFkQpb_LbFMthw.png" /></figure><p>Couple that with the fact that all of the items in the Archero pass are obtainable through free-to-play means, and you have a recipe for quite a disappointing purchase. There are no exclusive cosmetics here, and you can bet that Archero’s pass next season will look very similar to the current season. The rewards from this pass are ephemeral consumables, whereas in Fortnite’s pass the rewards are unique and permanent. Players should feel like they have received something special for buying the pass, instead Archero’s pass is peddling convenience items minus the convenience (you still need to grind through the pass to unlock them). The player’s perceived utility of the rewards included in these consumables-based battle passes decreases over time, as they progress further and the number of consumables required to upgrade their character/buildings/weapons also increases. So the battle pass becomes less and less of a worthwhile option for them. This is important because it requires you to think a few months ahead, maybe 12 months or longer, whether or not your battle pass will still offer value to your players. If the answer is “No” then you will need another IAP which is just as valuable to players on Day180 as it is on Day1. Cosmetics are almost completely invulnerable to this ‘shrinking utility’ since they aren’t used to progress, and no two cosmetics are the same. See below for a hypothetical example of consumables utility decay in a character upgrade economy, pay close attention to the green line:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*fsI-2-PwTexe-XJjb8OfNQ.png" /><figcaption><a href="https://dscarter95.github.io/medium_data_viz/utility_decay.html">https://dscarter95.github.io/medium_data_viz/utility_decay.html</a></figcaption></figure><p>In my opinion this is a really significant difference between the two passes, but a player’s interpretation of reward value (especially cosmetics) is something that’s extremely difficult to measure. Even so, I’m in no doubt that the conversion rate of the Archero pass is significantly lower as a result. This is something I saw on a previous project with a gacha-based character upgrade economy: late-game conversion rate was lower than early game.</p><p>In the Fortnite Pass the skins and cosmetics themselves also act as an incentive for conversions. Players are reminded of this persistently: simply wearing a skin turns every player into a walking billboard for the primary form of in-game monetisation. Before each game begins there is a short ‘lobby time’ where players explore a much smaller island, which forces them close to the 99 other players, and crucially the skins they are wearing. It’s common to see players show off their dances, emotes, or building skills, and gather round other players to admire their cool rare cosmetics. This short waiting time is a 30 second ad of sorts for the in-game cosmetics, many of which are only available in the battle pass for a <em>limited time</em>, remember.</p><p>While Archero does have a multiplayer mode, it doesn’t have such strong social elements baked into the game, so the opportunities to show off cosmetics are greatly reduced. The duos mode is only made available after completing the first 3 chapters, which can take some time and the reality is that most players would have churned by that point. It’s also limited to one other player that you share your lobby code with, which reduces opportunities to show off or to see new cosmetics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xlWUKH3JSAOrWRfvYl3mXQ.png" /></figure><p>Another social element in the Fortnite pass which boosts conversions is the Party Assist. This allows party members to contribute towards completing your challenges which grant substantial battle pass XP. The social element at play here is reciprocity: it benefits both players to buy the battle pass, play together and progress together. Archero’s duos mode is roughly equivalent, offering gems, gold and XP for playing with friends. But Archero makes a point of stating that enemies killed in duos mode <em>do not contribute</em> to battle pass progress. This robs the Archero pass of the opportunity to maximise not only conversions, but retention and engagement too, by deliberately removing incentives for social play which could have synergised well with the battle pass.</p><p>Finally, Fortnite maximises conversions by coupling battle pass releases to their major update cadence. Every Season is paired with a new battle pass, and at the start of every new Season there is a huge update with content galore. It has built a reputation for itself with outstanding production quality wow-factor events that draw in the crowds, players new and old, just as the new battle pass launches. A prime example would be the Chapter 2 Black Hole event which made headlines globally. The timing of the events is no coincidence, and <a href="https://www.gamesindustry.biz/articles/2018-05-25-fortnite-sold-5-million-battle-passes-on-the-first-day-of-season-3">drives massive numbers of conversions for the battle pass</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/364/0*6U1dkh_ls8voyznM" /></figure><p>The marquee events and the battle pass feed back into each other. One would be less effective without the other, and this touches on the most significant drawback of the battle pass/update pairing: it is an <em>extremely</em> content hungry cycle. If you don’t have the LiveOps and Art capabilities to keep churning out huge updates with new content then your battle pass won’t hit it’s potential. As a result Archero doesn’t have such a strong LiveOps cycle compared to Fortnite (of course it doesn’t, who does?), but it could and should improve on its LiveOps offerings, and time these with new battle pass releases for maximum effect.</p><h4>Boosting Retention and Engagement:</h4><p>During the ~70 day wait in between major updates in Fortnite, it’s the job of the battle pass to keep those players around. It does so in two ways:</p><ol><li>Retention — players want to extract as much value as possible from their purchase, so keep playing to progress through the pass to achieve this.</li><li>Engagement — players return to play more often: to play with friends and give them XP boosts, to claim the regular battle pass milestone rewards, due to daily and weekly challenges boosting battle pass progression.</li></ol><p>The most obvious thing to mention is that purchasing the battle pass itself has a retention effect, similar to a sunk-cost fallacy. Users who make a purchase retain about 2 to 3 times longer than their non-spending counterparts. These users are encouraged to retain, and play to acquire as many of the cosmetics/items that they have paid to unlock, extracting <a href="https://www.researchgate.net/publication/267942356_Gamification_from_the_perspective_of_service_marketing">as much value as possible</a> from their purchase. Fortnite reinforces and enables this thinking — by including a total of 1500 VBucks within the battle pass, enough to recover the initial cost of 950 VBucks, plus some more to spend in the Item Shop. The Archero pass has no equivalent.</p><p>In Fortnite it’s only required that the player reach roughly level 60 in the battle pass to have earned back 1000 VBucks. This creates an unspoken ‘VBucks Profit’ rule of the Fortnite battle pass: it’s possible to make a single purchase of 1000 VBucks ($10) and use this to buy every subsequent battle pass, just by playing regularly and putting the earned VBucks into next season’s pass. This makes a first time purchase extremely tempting. It also works as both a conversion gimmick and retention incentive: once you’ve made the purchase, you’re even more determined to play and exploit this fact as much as you can! Few players manage this in reality, and more often than not they’ll spend the VBucks they earned before the next pass is released.</p><p>Cosmetics and rewards are laid out along the Fortnite pass in such a way as to maximise engagement and retention, similar to pacing in a typical story-driven game. Notice in the plot below the <a href="https://www.gamasutra.com/view/feature/3848/gameplay_fundamentals_revisited_.php?print=1">regularly spaced peaks and troughs</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*_8zVlCiKKL_6a-URuH2W8g.png" /><figcaption><a href="https://dscarter95.github.io/medium_data_viz/fortnite-free-paid.html">https://dscarter95.github.io/medium_data_viz/fortnite-free-paid.html</a></figcaption></figure><p>This places regular high-value rewards just on the horizon, encouraging the player to keep going. This isn’t just for spenders either, but the free tier of the Fortnite battle pass also spaces out rewards at regular intervals. Where the paid tier attempts to keep players incentivised and engaged with a rollercoaster of high and low value rewards, the free tier instead focuses rewards much earlier on. The first free rewards occur frequently, gradually decreasing in frequency and then also decreasing in value beyond level 53. By this point the battle pass is focused on converting these users, and rightly so; conversion rate in the days following D7 is about 30% higher than before D7. Couple this with the nature of the battle pass increasing in value the more a player plays, and you have a perfect recipe for boosting conversion rate, and retention. We can thus describe the Fortnite battle pass in two phases for free-to-play users:</p><ol><li>Engagement focus up to level 53.</li><li>Conversion focus from then on.</li></ol><p>Archero’s battle pass follows a similar pattern of reward peaks and troughs, but with fewer levels, and instead opts to reward free players with a continuous stream of low-value rewards. Overall, free users can acquire roughly $15 worth of gifts from the Archero pass, whereas the Fortnite pass rewards free players with $64 worth. Keep in mind that these passes last for different durations, so Archero actually offers free players roughly $1 of stuff per day, whereas Fortnite offers $0.90 per day. Still, the Fortnite offerings feel more significant since they are lasting and unique compared to the consumables Archero offers.</p><figure><img alt="Hello" src="https://cdn-images-1.medium.com/max/700/1*SvgVFZNPxdMTh4ISbnX9pw.png" /><figcaption><a href="https://dscarter95.github.io/medium_data_viz/archero-free-paid.html">https://dscarter95.github.io/medium_data_viz/archero-free-paid.html</a></figcaption></figure><p>The Fortnite daily assignments and weekly challenges act in a similar way, providing measurable goals that grant XP and help the player progress to the next big reward in the battle pass. New assignments are given daily, providing short-term goals for players to focus on, and can include almost any in game action such as: eating apples 3 times in a specific location, or eliminating players in different areas of the map. Challenges provide mid-term goals and are released in stages throughout the season. A key point here is that many of these challenges aren’t related to skill, example: you don’t need to be great at FPS games to land at Lonely Lodge and do a dance, this makes battle pass progress accessible to all. To make it even more accessible Fortnite offers the ‘Party Assist’ function. Rewarding social activity is an effective retention strategy: if all of your friends are playing and helping each other out, it becomes more enjoyable for you to play too! Including social elements is a powerful KPI boosting tool in any Designer’s toolbox, especially when social interactions aid progress. For these social features to be impactful they must be embedded into the game design from the beginning, and <a href="https://jyx.jyu.fi/bitstream/handle/123456789/66822/URN%3ANBN%3Afi%3Ajyu-201912165314.pdf?sequence=1&amp;isAllowed=y">not shoehorned in</a> at a later date.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Wc-2qMGfLCSetmj-" /></figure><h3>Drawbacks and Pitfalls:</h3><p>As mentioned earlier, probably the most significant drawback of the battle pass monetisation system is how insanely content-hungry it is. You (or your studio) simply may not have the resources necessary to create and maintain a successful implementation like this. It may not be feasible for you to invest the time and resources into significant social features and events, which are signatures of a successful cosmetics driven battle pass. Even major studios often don’t have the resources to run LiveOps, and that’s before considering the additional art resources needed for exclusive cosmetics. As a result, you may be tempted to go down the route of tweaking the battle pass into a subscription-style IAP which provides bonuses akin to a VIP mechanic, like boosters and extra gacha pulls. Be warned, this approach often leads to cannibalisation and a reduction in revenue, and does not provide long term utility as discussed earlier. Remember:</p><blockquote>In order to make your battle pass as successful as possible, it should include items exclusive to that battle pass</blockquote><p>The economies in these games were not built around a battle pass, but an extended character/item gacha grind. While this design decision is not a problem in itself, it is almost completely incompatible with a battle pass which offers the same rewards that the gacha does. The Archero battle pass (and others like it) are focused on less compelling convenience items, limited or no cosmetics, and limited or no ‘social rewards’. By implementing a battle pass which gives players more of the same for a lower price, you will only cannibalise your other IAPs, which is exactly what has happened to several top-grossing mobile games that have tried this. In the best case you will see no increase in revenues with a poorly implemented pass, and a decrease in revenue in the worst case. But when done right <a href="https://www.pocketgamer.biz/news/72313/sensor-tower-clash-of-clash-727-million-usd-revenue-2019/">some games have seen increases in revenues of 27%</a>. This puts into perspective how important it really is to understand the nuance of the monetisation methods you consider for your game, and whether or not a battle pass is appropriate for the kind of game you have built (or want to build). It also highlights how, when done properly, a battle pass can significantly boost revenues and player engagement.</p><p>Another significant drawback to be aware of is that with cosmetics based economies there is naturally limited depth and therefore lower spend potential from individuals (or ‘whales’). Battles Passes rely heavily on high conversion rate but have lower ARPPU, so depending on the genre or type of game you are making it might make more sense to disregard broad appeal and higher conversion rate — in favour of depth and focus on your most dedicated players and unlocking their spend potential. In a cosmetics economy your spend depth is mostly limited by your cosmetics output, putting a significant strain on your Art resources. In this scenario, even your most dedicated customers cannot buy more than the limited number of skins, so coupling a cosmetics based battle pass with another, deeper form of in-game monetisation would be a smart move. Remember, this should not have any overlap with your battle pass offering.</p><h3>Summary:</h3><p>Let’s condense all that we’ve learned during this deep dive:</p><ol><li>Offer amazing rewards for the very first level of your battle pass, make purchasers feel great right away.</li><li>Fill it with exclusive items and varied cosmetics, not available via any other means.</li><li>Offer unique items with each pass to introduce time exclusivity.</li><li>Offer crazy total value for the entire pass, to the tune of 30x to 50x, to maximise conversions.</li><li>Be generous with your free rewards, be even more generous with your paid ones.</li><li>Make it last, 30 days should really be your minimum to maximise value and retention.</li><li>Make your paid pass valuable to players at all stages, early game and end game.</li><li>Give players maximum opportunity to show off their cosmetics to friends and in matchmaking.</li><li>Reward social play by providing boosts when playing with paid tier friends. Make rewards reciprocal.</li><li>Tie new battle pass launches to your major LiveOps events.</li><li>Pace free and paid rewards to encourage engagement. Always give the player an implicit ‘next big objective’.</li><li>Focus big value free rewards early, to engage and retain F2P players, then taper free rewards off to encourage conversion later on.</li><li>Consider offering hard currency in the free and paid tiers.</li><li>Give the player explicit bite-sized goals with challenges and missions that contribute to battle pass progress.</li><li>Make challenges accessible, non-skill based.</li><li>Offer additional quests/challenges/missions as part of the paid tier to enable more progression, and squeeze more hours out of the pass.</li><li>Don’t make consumables the primary rewards.</li><li>Don’t cannibalise your other IAPs.</li><li>Consider pairing a battle pass with another deeper form of monetisation to lift the spend cap of your most dedicated players, such as separate cosmetics.</li></ol><p>In the end, a battle pass, when done right, should offer the player superb value and an enhanced experience when playing your game. To achieve this you should design your battle pass around your core game at an early stage, but when that is not possible it should be designed to complement your existing features for maximum effect. I hope this article has gone some way towards helping you understand how that could be done in your game! :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=74b0a4041eff" width="1" height="1" alt=""><hr><p><a href="https://medium.com/game-science/the-battle-pass-how-to-get-it-right-74b0a4041eff">The Battle Pass — How to get it right</a> was originally published in <a href="https://medium.com/game-science">Game Science</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>