Parallel AutoRegressive Models for Multi-Agent Combinatorial Optimization

2024-09-05

AI Summary

Multi-agent combinatorial optimization (MACO) faces significant challenges—including inefficient coordination, poor generalization, and high inference latency—due to its NP-hard nature and the complexity of modeling inter-agent collaboration. To address these, we propose PARCO, a parallel autoregressive framework that abandons conventional sequential decoding and enables agents to jointly construct high-quality solutions in parallel. Our key contributions are: (1) the first parallel pointer mechanism enabling synchronous decision-making across multiple agents; (2) a priority-driven conflict resolution module ensuring solution consistency and feasibility; and (3) a Transformer-based communication layer coupled with a multi-pointer decoder, trained end-to-end via reinforcement learning. Evaluated on multi-agent vehicle routing and scheduling benchmarks, PARCO achieves substantial improvements over state-of-the-art methods, reduces inference latency by 67%, and demonstrates strong generalization to unseen problem scales.

Technology Category

• Multiagent Systems: Distributed Problem Solving
• Planning, Routing, and Scheduling: Planning with Language Models
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Agentic search
• Economics, Online Markets and Human Computation: Incentives in network design for Web infrastructures and ecosystems

Abstract

Combinatorial optimization problems involving multiple agents are notoriously challenging due to their NP-hard nature and the necessity for effective agent coordination. Despite advancements in learning-based methods, existing approaches often face critical limitations, including suboptimal agent coordination, poor generalizability, and high computational latency. To address these issues, we propose Parallel AutoRegressive Combinatorial Optimization (PARCO), a reinforcement learning framework designed to construct high-quality solutions for multi-agent combinatorial tasks efficiently. To this end, PARCO integrates three key components: (1) transformer-based communication layers to enable effective agent collaboration during parallel solution construction, (2) a multiple pointer mechanism for low-latency, parallel agent decision-making, and (3) priority-based conflict handlers to resolve decision conflicts via learned priorities. We evaluate PARCO in multi-agent vehicle routing and scheduling problems where our approach outperforms state-of-the-art learning methods and demonstrates strong generalization ability and remarkable computational efficiency. Code available at: https://github.com/ai4co/parco.

Problem

• Enhances multi-agent coordination in optimization
• Reduces computational latency in decision-making
• Improves solution quality and generalizability

Innovation

• Parallel AutoRegressive Combinatorial Optimization
• Transformer-based communication layers
• Multiple pointer mechanism for decision-making