BusEnv: A Multi-agent Reinforcement Learning Environment and Benchmark for Urban Public Transportation

WESLEY DA SILVA E SILVA, RICARDO RIOS, RAFAEL DA COSTA FONSECA, Sabarikirishwaran Ponnambalam, Léa Cassé, Marcos Vinícius dos Santos Ferreira, Albert Bifet, Tatiane Nogueira Rios

Keywords: Multi-agent environment, Reinforcement Learning, Public Transportation

TL;DR: BusEnv is a new MARL benchmark for urban transit using large-scale real data. It simulates bus agents balancing service quality and efficiency, provides baselines, and supports data-driven transit management.

Abstract: Reinforcement learning (RL) offers a powerful paradigm for managing complex, dynamic transportation systems where autonomous agents must adapt to uncertain and rapidly changing environments. We present BusEnv, a benchmark environment grounded in real-world data from the Salvador Urban Transportation Network, encompassing approximately 700,000 passengers, 2,000 vehicles, 400 lines, and 3,000 stops, collected between March 2024 and March 2025 at sub-minute resolution. BusEnv simulates realistic bus operations with stochastic passenger demand, route-specific travel times, and traffic-dependent variability, enabling controlled experimentation under partially observable, high-dimensional conditions. The reward function integrates multiple objectives, such as passenger service quality, operational efficiency, maintenance adherence, and sustainability, allowing the assessment of how different RL algorithms balance these competing factors. We evaluate nine baseline methods implemented in MARLlib, analyzing their convergence, robustness, and environmental impact when deployed under independent-learning conditions. Results show that PPO-based approaches achieve the highest stability and lowest energy waste, linking algorithmic robustness to sustainability performance. By combining data realism with reproducibility and extensibility, BusEnv establishes a foundation for systematic research on learning-based transport management and provides a scalable testbed for future studies on cooperative, sustainability-aware reinforcement learning.

Area: Engineering and Analysis of Multiagent Systems (EMAS)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 907