Multi-Goal Pathfinding with Deep Q-Learning

Pathfinding, the process of finding a traversable route from one point to another, is an important field of research due to the implications it can have for many other fields and technologies. The purpose of this paper is to propose a new Q-Learning based pathfinding algorithm to solve mazes in which the algorithm (“agent”) must find multiple subgoals before reaching a final destination, in fewer iterations than existing Q-Learning algorithms. The existence of multiple subgoals can result in ‘reward loops’ which cause an unnecessary increase in the number of iterations to solve the maze [1]. These reward loops can be eliminated with the use of Multiple Q-Tables [1]. However, Q-Table based methods require many iterations for large state environments (such as large mazes) to learn the Q-values for each state-action pair [2], [3]. Deep Q-Learning can be used to approximate Q-values for previously unseen state-action pairs and thus reduce the number of iterations to learn the optimal solution [3]. Therefore, we propose an algorithm that adapts Multiple Q-Table theory to Deep Q-Learning to address the two weaknesses of basic Q-Learning at once: 1) The poor scalability to large mazes, and 2) The existence of reward loops. The proposed design is the use of Multiple Deep Q-Networks, each of which is responsible for finding the shortest path to the nearest subgoal or final destination (from the previous subgoal or starting point). We hypothesize that in addition to the elimination of reward loops, this results in other advantages including the balancing of exploration and of the replay memory allocation. In addition, we optimize our design with an improved Exploration Strategy, the addition of a Revisiting Penalty, as well as hyperparameter optimization. We test our solution on sample mazes of four sizes and compare it to the Multiple Q-Table and Single Deep Q-Network algorithms. Our results confirm our hypothesis and show that our solution outperforms the other algorithms in the number of iterations to find the shortest path, especially on larger mazes. Finally, we offer suggestions for alternative designs, future work, and improvements.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.