pyagents is a small collection of tools and agents for a university project: Make an agent that learns to play Pong, the Atari 2600 game.

• It is assumed that ale_python_interface (ALE) is installed.

• It is assumed that there is a directory on the same level as pyagents containing Atari ROMs, which must adhere to the naming standard defined by ALE.

• game_manager.GameManager handles the interface between our agents, ALE, and a logging system.

• The agents subpackage contains agents that all fulfill the interface defined by agents.Agent.

• The experiments subpackage contains expirement recipes, that when imported by run_experiment.py conduct a full experiment with a GameManager and an agent.

• You conduct an experiment by e.g. ./run_experiment.py sarsa2.

• evaluate_experiment.py is used like run_experiment.py, and provides automated experiment evaluation.

• Pong-specific preprocessing tools are found in util.pongcess

Do note that the performance is not impressive. We share the code in the hope that it may help others by providing ideas for a way of structuring the task, and by showing our take on the implementation on different tools such as CircularList.

If you would like to see an approach to dynamic module loading, check out experiments/__init__.py and its usage in run_experiment.py and evaluate_experiment.py.