This is a python (tensorflow) implementation of the following paper in python: https://arxiv.org/pdf/1511.08779.pdf The goal is to retain the ball on the table as long as possible. I have implemented it by writing a custom pong environment because the Atari one seems to be available only for Lua.
This is a collection of samples taken at lengthly intervals, training on my CPU:
Just look at this now! (recorded without delay)
There are two different networks (which consist of Convolutional, Dense and GRU layers) controlling the two paddles. They are made to cooperate by giving a negative reward to both when one of them leaves the ball. When we give different rewards, for instance x for the winning one and -1 for the losing one, by changing x from -1 to 1, we transition the system from coordination to competition (as given in the paper). In order to experiment with such rewards, simply change the rewards given in the Pong class.
Simply run train.py
The above code automatically configures Tensorboard. Tensorboard has been configured to bind with port 8010. If yours isnt free, change it in config.py. The console will print out a URL (local), simply open up that in your browser to check the progress. Tensorboard will show only one plot: the average number of hits over the last 30 episodes.