CLEAR: A Congestion-Limited, Event-Aware Reinforcement Learning Solution for Preventing Cooperation Collapse in Multi-Agent Systems
ISEF · 2026 Software Design
Overview
Multi-agent AI systems are increasingly prevalent in various fields ranging from autonomous drone systems to healthcare assistant robots. Achieving success in complex environments requires effective communication protocols for AI agents. This project investigates the non-monotonic relationship between cooperation and utility maximization while proposing a novel congestion-limited, event-aware (CLEAR) solution for communication failure. Cooperation collapse in multi-agent reinforcement learning (MARL) occurs when constraints including limited bandwidth, transmission costs, congestion, and receiver overload impair communication protocols. A novel MARL pipeline that incorporates a receiver-aware speaker agent, stochastic delivery channel, mid-episode goal switch, and belief latch is developed. Proximal policy optimization is implemented to update each agent’s parameters. By modeling real-world constraints and using these parameters to probabilistically adjust learned protocols, communication is controlled through a gating policy. The multi-agent system is trained in 300 training episodes at 10 different seeds for each experimental configuration. Experimental results demonstrate non-monotonic behavior with larger budgets in real-world settings, validating the cooperation collapse phenomena proposed by this project. When testing the novel CLEAR algorithm against baseline solutions, CLEAR demonstrates statistically significant improvement in the success per message rate by an average of 53.6%, emphasizing the role of the new algorithm for the future of multi-agent reinforcement learning.
Awards (2)
- Fourth Award of $600 $600
- Association for the Advancement of Artificial Intelligence: AAAI Student Memberships for each finalist that is part of the 1st, 2nd, and 3rd Prize Winning projects and 5 Honorable Mention winning projects (up to 3 students per project) (in-kind award / part of the 1st-3rd prize)
Competition history
- ISEF 2026
Resources
Related projects
ISEF · 2020
Mission-Critical Communications Planning Over Contested RF Spectrum with Deep Reinforcement Learning Aided Artificial Intelligence
ISEF · 2019
A Game of Jamming: A Multi-Agent Game Theoretic Learning Based Cognitive Anti-Jamming Communication System to Combat an AI Jammer
ISEF · 2026
Meta-Reinforcement Learning Framework for Network-Level Bridge Management
ISEF · 2019
Novel Reinforcement Learning Methods in Collaborative Environments
Closest projects by meaning, across every fair and year in the corpus.
Source: Regeneron International Science and Engineering Fair