Path shape generation control in a maze by using curriculum learning

Abstract

In this study, PCGRL(Procedural Contents Generation via Reinforcement Learning) controls the path shape of the maze. In addition, we propose a generative method that applies curriculum learning and report on the results. In this case, the goal of the control is to make the shape of the shortest path of the maze in spiral. This is because we expected the longest path takes a spiral shape within a certain area and conditions. The details of the proposed method are as follows: the reward function is set so that the closer the shortest path is to a spiral shape, the more reward is given, and the goal position of the maze is moved spirally according to the degree of task accomplishment. As a result, the reward value was stably incremented in curriculum learning, but the difficulty and conditions for task updating remain an issue.