Believable self-learning AI for world of tennis

Abstract

We describe a method used to build a practical AI system for a mobile game of tennis. The chosen approach had to support two goals: (1) provide a large number of believable and diverse AI characters, and (2) let the users train AI “ghost” characters able to substitute them. We achieve these goals by learning AI agents from collected behavior data of human-controlled characters. The acquired knowledge is used by a case-based reasoning algorithm to perform human-like decision making. Our experiments show that the resulting agents indeed exhibit a variety of recognizable play styles, resembling the play styles of their human trainers. The resulting AI system demonstrated stable decision making, adequate for use in a real commercial game project.