Tutorial 16, added note about compatibility issues.

Author: Hvass-Labs

16_Reinforcement_Learning.ipynb

@@ -533,6 +533,10 @@
     "\n",
     "You can download a TensorFlow checkpoint which holds all the pre-trained variables for the Neural Network. Two checkpoints are provided, one for Breakout and one for Space Invaders. They were both trained for about 150 hours on a laptop with 2.6 GHz CPU and a GTX 1070 GPU.\n",
     "\n",
+    "#### COMPATIBILITY ISSUES\n",
+    "\n",
+    "These TensorFlow checkpoints were developed with OpenAI gym v. 0.8.1 and atari-py v. 0.0.19 which had unused / redundant actions as noted above. There appears to have been a change in the gym API since then, as the unused actions are no longer present. This means the vectors with actions and Q-values now only contain 4 elements instead of the 6 shown here. This also means that the TensorFlow checkpoints cannot be used with newer versions of gym and atari-py, so in order to use these pre-trained checkpoints you need to install the older versions of gym and atari-py - or you can just train a new model yourself so you get a new TensorFlow checkpoint.\n",
+    "\n",
     "#### WARNING!\n",
     "\n",
     "These checkpoints are 280-360 MB each. They are currently hosted on the webserver I use for [www.hvass-labs.org](www.hvass-labs.org) because it is awkward to automatically download large files on Google Drive. To lower the traffic on my webserver, this line has been commented out, so you have to activate it manually. You are welcome to download it, I just don't want it to download automatically for everyone who only  wants to run this Notebook briefly."
@@ -4874,7 +4878,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.6.1"
+   "version": "3.6.0"
   }
  },
  "nbformat": 4,