A curriculum-based reinforcement learninig approach to pedestrian simulation

Reinforcement Learning represents a way to train an agent situated in an environment what to do to maximise an accumulated numerical reward signal (received by the environment as a feedback to every chosen action). Within this paper we explore the possibility to apply this approach to pedestrian modelling: pedestrians generally do not exhibit an optimal behaviour, therefore we carefully defined a reward function (combining contributions related to proxemics, goal orientation, basic wayfinding considerations), but also a particular training curriculum, a set of scenarios of growing difficulty supporting the incremental acquisition of proper orientation, walking, and pedestrian interaction competences. The paper will describe the fundamental elements of the approach, its implementation within a software framework employing Unity and ML-Agents, describing the promising achieved simulation results.