TSN-Affinity: Parameter Reuse for Continual Offline RL

Researchers propose TSN-Affinity, a method for continual offline reinforcement learning (CORL) that uses TinySubNetworks and Decision Transformer to enable task-specific parameter reuse. CORL involves learning from static datasets across sequential tasks while avoiding catastrophic forgetting. Replay-based methods suffer from memory overhead and distribution mismatch, while architectural approaches remain underexplored in CORL. TSN-Affinity addresses these challenges by leveraging similarity-driven parameter reuse.