Generative Exploration for Offline Reinforcement Learning

Codebase: In Progress

Abstract: We consider an offline reinforcement learning scenario where the data is limited. Decision transformers are a type of transformer used in reinforcement learning that maps trajectory context to determine the next action. One special property of decision transformers is its ability to generalize to producing viable actions for trajectories it has not seen before. Using this property, we hope to synthetically create offline reinforcement learning data that can be used to train models in scenarios where data is sparse.

Paper: In Progress