Cooperative, hybrid multi-agent systems for distributed, real-time traffic signal control

Abstract

This study presents several innovative distributed cooperative problem solving approaches and continuous online learning methodologies by incorporating advanced cooperation mechanisms as well as synergistic hybrid computational intelligence techniques in intelligent multi-agent systems. Three autonomous multi-agent systems have been developed to solve the complex problem of providing real-time traffic signal control for a large simulated real world traffic network. The first hybrid multi-agent system has a hierarchical architecture, in which each agent uses a multi-stage online learning process to adapt itself to the changing problem. The second hybrid multi-agent system is designed mainly by combining concepts from stochastic approximation theorems, neural networks and fuzzy logic. The final multi-agent system developed in this study is capable of allocating cooperative zones dynamically thus allowing dynamic team-building. The three multi-agent systems have shown to perform better than GLIDE, a simulated version of the current traffic signal control system used in Singapore, in all of the simulation scenarios. The positive results from this study presents exciting new directions for developing multi-agent systems with advanced cognitive capabilities for solving real world problems.