A discussion on planning method in platform game AI

–A study of situation recognition for switching plans–

Abstract

Planning is important issue for autonomous agent. Though many researchers proposed effective planning methods, they are not sufficient. We proposed the planning method that switches Q-learning that is a kind of reinforcement learning methods and A* algorithm that is a kind of search methods. In this method, three algorithms are used: the plan which used A* algorithm and two plans which used Q-learning. In this article, we aim to improve the method in computer game "Infinite Tux" which is a kind of benchmark. It decide action by switching plans according to state around agent. We discuss the impact on efficiency of planning by changing the number of plans in this method.