In grid-world settings, LYCO consistently outperformed representative decentralized RL baselines in high-density environments, achieving a success rate above 0.7 at an agent density of 0.3 on a $20\times 20$ map with obstacle density 0.1 while maintaining a competitive makespan.
Abstract
Leveraging its flexibility and scalability, decentralized reinforcement learning (RL) has been widely applied to multi-agent path finding (MAPF). However, existing methods often exhibit a decline in success rate as agent density increases. For example, agents that reach their goals early may remain there and block other agents. To maintain a high success rate as the number of agents grows, we propose Learning to Yield for Decentralized MAPF in Congested Environments (LYCO), a decentralized policy that explicitly learns localized yielding behaviors. LYCO is trained by imitation learning from Priority Inheritance with Backtracking (PIBT), a lightweight heuristic solver, and adds two components that promote yielding: (i) an observation channel encoding the estimated next positions of higher-priority neighboring agents, and (ii) a local cooperative reward that discourages waiting at a goal unless nearby agents have also reached theirs. Crucially, PIBT serves only as a source of imitation targets during training; during both RL training and inference, each agent selects actions solely from its local observations using the learned policy. LYCO thus maintains decentralized action selection without performing priority inheritance or backtracking at inference time. In grid-world settings, LYCO consistently outperformed representative decentralized RL baselines in high-density environments, achieving a success rate above 0.7 at an agent density of 0.3 on a $20\times 20$ map with obstacle density 0.1 while maintaining a competitive makespan. These findings show that LYCO is an effective decentralized approach for congested MAPF.
This work instantiates MAPL using MARL from human feedback with a learned reward model and multi-agent direct preference optimization (MADPO), and shows that MAPL can improve collaboration quality and efficiency while approaching the performance of MARL with fixed, well-defined rewards.
ICMAPE converts the fixed-confidence identification objective into a reward derived from inference confidence, so that standard reinforcement learning machinery can be applied to decentralized pure exploration.
G2MAF applies one globally normalized, projected critic gradient to guide and coordinate all agents'corrections while keeping the action both feasible and close to the frozen policy proposal.
Guo-Wei Zou, Hai-Tao Wang, Guo-Xin Wang et al.· arXiv.org· 0 citations
In cooperative Multi-Agent Reinforcement Learning (MARL), agents must learn to coordinate their actions to achieve shared objectives. A key indicator of effective coordination is behavioral diversity, i.e., the agents'potential to enact distinct yet complementary control policies given the same set of observations. Alt...
This study develops CoRe-MARL, a cooperative multi-agent reinforcement learning (MARL) framework, by formulating a decentralized partially observable Markov decision process (Dec-POMDP), and finds that MAPPO reduces the service gap across local centers and enhances service for the worst-served center while maintaining...
Naimur Rahman Chowdhury, Shatabdi Sen Prapti, Md. Salehin Seyam et al.· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.