kdh-yu/Reinforcement-Learning-Comparative-Study

Comparative study of Reinforcement Learning Algorithms on Ping Pong game: In the current design of experience replay we sample uniformly to obtain the minibatch and update the model. Devising a way to sample more experience points close to the tricky areas would help solving this problem, better the training rate and improve convergence. We designed a game environment for the Android platform as few such environments are available at the moment. Moreover, during the pre-processing game we removed the background and score to reduce clutter and increase likeliness of successful training. It would be interesting to see how restoring the background affects agent’s performance. Overall, our results show the capacity of Deep neural networks and how a generic reinforcement learning setup such as this could learn and play the game with very minimal domain knowledge.

Jupyter NotebookMIT