Automatic parameter selection for multi-speedmulti-agent pathfinding solver

Abstract

Multi-agent pathfinding is a study focused on finding collision-free paths for multiple agents in a shared environment in order for the agents to reach designated destinations. As the problem is well researched, more and more sophisticated algorithms are available that utilize different parameters to fine-tune the search for these solutions. In this paper, we demonstrate the combining of multi-agent pathfinding with machine learning to perform such parameter selection dynamically. This selection is approached as an image classification problem, where unseen problem instances are divided between a collection of available parameter values. For this, typical vehicular road traffic scenarios were implemented for experimentation. The presented results give new knowledge for this previously unstudied approach indicating that there are cases in which visual representation of the problem is sufficient as the dynamic selection can perform better when compared to a fixed setting. Therefore, demonstrating the potential that machine learning can be used to improve the performance of a multi-agent pathfinding algorithm without needing to develop it further.