Conservative Reward-Action Balancing Transformer

Research in goal-conditioned reinforcement learning (GCRL) aims to deploy trained agents in realistic settings. Offline reinforcement learning has gained attention as a method to minimize the costs associated with online interactions in GCRL. One approach, the Decision Transformer (DT), leverages a numerical target known as ‘return-to-go’ to achieve enhanced performance. However, because DT assumes an ideal environment with complete knowledge of rewards, there is a need to develop improved techniques for real-world scenarios. This study explores various strategies and outcomes for conservative reward-action balancing transformers designed to function effectively under practical conditions.