| dc.description.abstract | This paper aims to investigate the potential of model-free reinforcement learning using the Tsetlin Machine by evaluating its performance in widely recognized benchmark environments for reinforcement learning: Cartpole and Pong. Our study is divided into two primary objectives. First, we analyze the effectiveness of the Tsetlin Machine in learning from the actions of expert agents in the Cartpole environment. Second, we assess the ability of the multiclass Tsetlin Machine to learn to play both Cartpole and Pong environments from scratch.
Our findings indicate that the Tsetlin Machine can successfully learn and solve the Cartpole environment. Although the Pong environment remains unsolved, the Tsetlin Machine demonstrates its learning capabilities by scoring several points in multiple test runs, even managing to win in some of them. Through our empirical investigation, we conclude that the Tsetlin Machine exhibits promise in the field of reinforcement learning. Nonetheless, further research is needed to address the limitations observed in its performance in some of the examined environments. | |
