CS5180 Reinforcement Learning Final Northeastern University

For my final project in CS5180 Reinforcement Learning at Northeastern University I did a SARSA(on-policy) vs. Q-learning(off-policy) and SARSA single-agent vs. batch/genetic learning comparison on performance of tabular learning algorithms on Tetris called TetrisRL

This at a high level involved creating a custom py-game client, based on Tetris Py-game tutorial from Net Ninja, with some improvements/modifications. Using this and a custom threaded learning approach to have the agent interact with the game, multiple models were trained for the above comparisons mentioned.

• SARSA(on-policy) vs. Q-learning(off-policy)

• SARSA single-agent vs. batch/genetic learning

To the right is an example training sequence, and below an embedded copy of the final presentation with more detail on the project.