Continual Knowledge Adaptation for Reinforcement Learning

Bin Xiao (Chongqing University of Posts and Tel.) · Jinwu Hu (South China University of Technology) · ZiHao Lian (South China University of Technology) · Zhiquan Wen (South China University of Technology) · Chenghao Li (South China University of Technology) · Guohao Chen (South China University of Technology) · Xutao Wen (South China University of Technology) · Mingkui Tan (South China University of Technology)

adaptive knowledge mergingbenchmark experiments catastrophic forgetting continual learning forward transfer knowledge adaptation knowledge transferknowledge utilizationknowledge vector poolmemory requirements model parameters performance improvement reinforcement learning scalability challenges task-specific knowledge

Reinforcement Learning enables agents to learn optimal behaviors through interactions with environments. However, real-world environments are typically non-stationary, requiring agents to continuously adapt to new tasks and changing conditions. Although Continual Reinforcement Learning facilitates learning across multiple tasks, existing methods often suffer from catastrophic forgetting and inefficient knowledge utilization. To address these challenges, we propose Continual Knowledge Adaptation for Reinforcement Learning (CKA-RL), which enables the accumulation and effective utilization of historical knowledge. Specifically, we introduce a Continual Knowledge Adaptation strategy, which involves maintaining a task-specific knowledge vector pool and dynamically using historical knowledge to adapt the agent to new tasks. This process mitigates catastrophic forgetting and enables efficient knowledge transfer across tasks by preserving and adapting critical model parameters. Additionally, we propose an Adaptive Knowledge Merging mechanism that combines similar knowledge vectors to address scalability challenges, reducing memory requirements while ensuring the retention of essential knowledge. Experiments on three benchmarks demonstrate that the proposed CKA-RL outperforms state-of-the-art methods, achieving an improvement of 4.20% in overall performance and 8.02% in forward transfer. The source code is available at https://github.com/Fhujinwu/CKA-RL.