Plasticity from Structured Sparsity: Mastering Continual Reinforcement Learning through Fine-grained Network Allocation and Dormant Neuron Exploration
Keywords: Continual reinforcement learning, Policy transfer
TL;DR: SSDE improves Continual Reinforcement Learning by balancing plasticity and stability to prevent forgetting.
Abstract: Continual reinforcement learning faces a central challenge in striking a balance between plasticity and stability to mitigate catastrophic forgetting. In this paper, we introduce SSDE, a novel structure-based method that aims to improve plasticity through a fine-grained allocation strategy with Structured Sparsity and Dormant-guided Exploration. Specifically, SSDE decomposes the parameter space for each task into forward-transfer (frozen) parameters and task-specific (trainable) parameters. Crucially, these parameters are allocated by an efficient co-allocation scheme under sparse coding, ensuring sufficient trainable capacity for new tasks while promoting efficient forward transfer through frozen parameters. Furthermore, structure-based methods often suffer from rigidity due to the accumulation of non-trainable parameters, hindering exploration. To overcome this, we propose a novel exploration technique based on sensitivity-guided dormant neurons, which systematically identifies and resets insensitive parameters. Our comprehensive experiments demonstrate that SSDE outperforms current state-of-the-art methods and achieves a superior success rate of $95\%$% on CW10 Continual World benchmark.
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13768
Loading