TSN-Affinity: Similarity-Driven Parameter Reuse Solves Catastrophic Forgetting in Continual Offline RL

arXiv·medium signal

Continual offline reinforcement learning (CORL) requires learning new tasks from sequentially arriving datasets without forgetting prior tasks — a practical scenario for any deployed RL system that adapts over time. TSN-Affinity uses task-similarity metrics to decide which previous network parameters to reuse vs. reinitialize for new tasks, avoiding both catastrophic forgetting and negative transfer. Benchmark results show significant improvements over existing CORL methods. Relevant to practitioners deploying adaptive RL agents in production environments with evolving task distributions.