Towards multi-agent reinforcement learning for integrated network of optimal traffic controllers (MARLIN-OTC)

Traffic congestion can be alleviated by infrastructure expansions; however, improving the existing infrastructure using traffic control is more plausible due to the obvious financial resources and physical space constraints. The most promising control tools include ramp metering, variable message signs, and signalized intersections. Synergizing the aforementioned strategies in one platform is an ultimate and challenging goal to alleviate traffic gridlock and optimally utilize the existing system capacity; this is referred to as Integrated Traffic Control (ITC). Reinforcement Learning (RL) techniques have the potential to tackle the optimal traffic control problem. Game Theory (GT) fits well in modelling the distributed control systems as multiplayer games. Multi-Agent Reinforcement Learning (MARL) achieves the potential synergy of RL and GT concepts, providing a promising tool for optimal distributed traffic control. The objective of this paper is to clarify the opportunities of game theory concepts and MARL approaches in creating an adaptive optimal traffic control system that is decentralized but yet integrated through agents' interactions. In this paper, we comparatively review and evaluate the relevant existing approaches. We then envision and introduce a novel framework that combines GT concepts and MARL to achieve a Multi-Agent Reinforcement Learning for Integrated Network of Optimal Traffic Controllers (MARLIN-OTC).