Deep RL
Deep reinforcement learning (Deep RL) aims to train agents to make optimal decisions in complex environments by learning from experience, primarily through trial and error. Current research emphasizes improving sample efficiency, addressing challenges like value overestimation and hyperparameter sensitivity, and enhancing the interpretability and robustness of learned policies across diverse domains. Prominent algorithms include actor-critic methods (e.g., A2C, PPO, SAC, TD3, DDPG), and research explores architectures like Mixture-of-Experts networks and the integration of symbolic reasoning and program synthesis for improved generalization and long-horizon task solving. These advancements hold significant potential for applications in robotics, finance, autonomous driving, and other fields requiring adaptive decision-making in dynamic environments.
Papers
On the consistency of hyper-parameter selection in value-based deep reinforcement learning
Johan Obando-Ceron, João G.M. Araújo, Aaron Courville, Pablo Samuel Castro
Performance Comparison of Deep RL Algorithms for Mixed Traffic Cooperative Lane-Changing
Xue Yao, Shengren Hou, Serge P. Hoogendoorn, Simeon C. Calvert