A decentralized multi-agent-based optimization framework for intelligent traffic systems

Abstract

The increasing complexity of urban transportation has increased the importance of intelligent traffic systems (ITS). Conventional approaches, such as fixed-time control (STS), rule-based adaptive control (RBAC), and centralized optimization (CTO), struggle with traffic demand, and network complexity. These limitations highlight the need for decentralized control strategies. To address this challenge, this study proposes a Decentralized Agent-Based Multi-Agent Reinforcement Learning Traffic Optimization (DAB-MARLTO) framework for smart traffic networks. The framework modeled urban road as a directed graph where vehicles, traffic signals, and road segments operate as autonomous agents. Traffic demand is modeled using a Poisson arrival process, while vehicle behavior is simulated via car-following, lane-changing, and queue models. Each signalized intersection is modeled as a Markov Decision Process (MDP) and optimized using decentralized tabular Q-learning (QL). The framework is implemented in Simulation of Urban Mobility (SUMO) via Traffic Control Interface (TraCI) policy execution. Performance metrics include Average Travel Time (ATT), throughput, average queue length, Congestion Index (CI), fuel consumption, and emission intensity. Results show DAB-MARLTO outperforms STS, RBAC, and CTO by reducing network delay, queue length, and congestion while improving efficiency. The framework achieves a 26.9% reduction in travel time, 46.4% reduction in queue length, 33.3% improvement in congestion index, and 27.4% increase in network throughput.