<ul><li>Regularization, data augmentation, and sparse reward functions are used in reinforcement learning to promote simplicity and increase generalizability.</li><li>A new approach is introduced to maximize the total correlation within induced trajectories in reinforcement learning, aiming to promote simple behavior throughout the episode.</li><li>An algorithm is proposed to optimize policy and state representation models based on a lower-bound approximation, resulting in superior robustness and improved performance in simulated robot environments.</li><li>The method naturally generates policies with periodic and compressible trajectories, showing better resistance to noise and dynamic changes compared to existing methods.</li></ul>

Maximum Total Correlation Reinforcement Learning

Discover more