Multi-Agent Path Finding using Dynamic Distributed Deep Learning Model

摘要

Multi-Agent Pathfinding (MAPF) is the task of finding paths for more than one agent in a common environment such that each agent reaches its goal and does not collide. Multi-Agent Path Finding is a problem that is related to various applications ranging from robot movements, game simulations, self-driving vehicles, etc. Most of the existing solutions to these problems are providing paths to each agent which they shall follow. These paths are optimal but are calculated once for a certain fixed and known scenario. There might be collisions and time-based overlapping paths if another agent or an obstacle is dynamically added to the grid. Thus, there is a need to have a system that deals with such situations while providing the most possible optimal path. Hence, we propose a solution to this problem by using a Deep Learning Model trained on a dataset made using an A* pathfinding algorithm that can take decisions at run time for each agent while making sure it does not collide with any obstacle or any other agent and handles all sort of dynamic variations in the environment at any point of time while not relying on precomputation.