BRIDGEBENCH: AN OFFLINE CONSTRAINED MULTI- AGENT REINFORCEMENT LEARNING BENCHMARK FOR INFRASTRUCTURE MANAGEMENT

ICLR 2026 Conference Submission18464 Authors

Keywords: Multi-Agent Reinforcement Learning, Offline Reinforcement Learning, Constrained Reinforcement Learning, Infrastructure Management, Bridge Maintenance, Benchmark, Real-world Applications, Decision Making

Abstract: Effective infrastructure management, particularly bridge maintenance, is critical for public safety and economic benefits. Reinforcement Learning (RL) offers a promising paradigm for optimizing maintenance policies. Real-world applications often involve multiple decision-makers, rely on pre-collected offline data, and necessitate strict adherence to operational constraints. Existing RL benchmarks and methodologies frequently fall short in simultaneously addressing these multi-agent, offline, and constrained aspects within a practical domain. To bridge this gap, we introduce BridgeBench, a novel offline constrained multi-agent RL benchmark for bridge maintenance. It provides a realistic and challenging environment for evaluating algorithms designed for complex infrastructure management tasks. We integrate various state-of-the-art single-agent and multi-agent offline constrained RL algorithms on this platform, providing insights into their performance and limitations. Our work aims to accelerate research in applying advanced RL techniques to critical real-world infrastructure challenges, fostering the development of more robust, safe, and cost-effective maintenance strategies.

Primary Area: datasets and benchmarks

Submission Number: 18464