Enhancing Cooperative Problem-Solving in Sparse-Reward Systems via Co-evolutionary Curriculum Learning
Keywords: Reinforcement Learning, Task Learning
Abstract: Sparse reward environments consistently challenge reinforcement learning, as agents often need to finish tasks before receiving any feedback, leading to limited incentive signals. This issue becomes even more pronounced in multi-agent systems (MAS), where a single reward must be distributed among multiple agents over time, frequently resulting in suboptimal or inconsistent learning outcomes. To tackle this challenge, we introduce a novel approach called Collaborative Multi-dimensional Course Learning (CCL) for multi-agent cooperation scenarios. CCL features three key innovations: (1) It establishes an adaptive curriculum framework tailored for MAS, refining intermediate tasks to individual agents to ensure balanced strategy development. (2) A novel variant evolution algorithm creates more detailed intermediate tasks. (3) Co-evolution between agents and their environment is modeled to enhance training stability under sparse reward conditions. In evaluations across five tasks within multi-particle environments (MPE) and Hide and Seek (Hns), CCL demonstrated superior performance, surpassing existing benchmarks and excelling in sparse reward settings.
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 13141
Loading