Interesting State Learning as a paradigm for Reinforcement Learning in Mario

In this project, we aim to train a RL agent to play Super Mario Bros for the Nintendo Entertainment System (NES). Taking inspiration from the intricate nature of human learning, we aim to identify trajectories with ”weak performance”, enabling the agent to engage in targeted ”practice” sessions, emulating the iterative learning process observed in human skill development through a new strategy called Interesting State Learning. Through this approach, we hope to reduce convergence times for the agent as well as enhance generalization to unseen levels.

Link to the report: Interesting State Learning as a paradigm for Reinforcement Learning in Mario