Can motorcyclist behavior in traffic conflicts be modeled? A deep reinforcement learning approach for motorcycle-pedestrian interactions

Abstract

Traffic simulation models have been utilized recently in safety evaluations by calculating traffic conflict indicators from simulated road user trajectories. However, existing simulation models (1) do not accurately capture road users’ behaviour and evasive actions, and (2) do not consider road users’ intelligence and rationality. This study proposes a safety-oriented microsimulation framework for modeling conflict interactions between motorcyclists and pedestrians in mixed traffic conditions. Motorcyclists are modeled as utility-based rational and intelligent decision-makers using a Markov Decision Process (MDP). Continuous Inverse Reinforcement Learning (IRL) is proposed to recover the motorcyclists’ reward function using their actual trajectories. The recovered motorcyclists’ reward function provides inferences into their behavior in conflict interactions. The motorcyclists’ optimal policies (sequences of decisions) are estimated using the Actor-Critic Deep Reinforcement Learning. The results show that the model simulated motorcyclist trajectories and the evasive actions with high accuracy. Moreover, the predicted PETs correlated well with corresponding field-measured conflicts.