Generation of Traffic Flows in Multi-Agent Traffic Simulation with Agent Behavior Model based on Deep Reinforcement Learning

TL;DR

This paper introduces a deep reinforcement learning approach for multi-agent traffic simulation, enabling agents to behave more realistically and recognize other agents' actions, resulting in more diverse and human-like traffic flows.

Contribution

It proposes a unified deep reinforcement learning framework combining visual and numerical data for realistic agent acceleration decisions in traffic simulation.

Findings

Agents recognize surrounding behavior effectively.

Traffic flows exhibit realistic diversity.

Agents learn to accelerate and decelerate irregularly.

Abstract

In multi-agent based traffic simulation, agents are always supposed to move following existing instructions, and mechanically and unnaturally imitate human behavior. The human drivers perform acceleration or deceleration irregularly all the time, which seems unnecessary in some conditions. For letting agents in traffic simulation behave more like humans and recognize other agents' behavior in complex conditions, we propose a unified mechanism for agents learn to decide various accelerations by using deep reinforcement learning based on a combination of regenerated visual images revealing some notable features, and numerical vectors containing some important data such as instantaneous speed. By handling batches of sequential data, agents are enabled to recognize surrounding agents' behavior and decide their own acceleration. In addition, we can generate a traffic flow behaving diversely to simulate the real traffic flow by using an architecture of fully decentralized training and fully centralized execution without violating Markov assumptions.