An adaptive game opponent

Abstract

Video games have been a popular source of entertainment for quite some time. A player wins a typical game by overcoming obstacles and non-player characters (NPCs) on the way to achieving a specific goal. With today's hardware, the virtual worlds which a game sets the player in look very realistic. However the opponents within most games do not behave realistically and appear unintelligent. Artificial Intelligence (AI) techniques have been used in games in an attempt to make NPCs seem more intelligent. Nevertheless, human players prefer to interact with other players rather than with computer-controlled ones in a game since even with AI, NPCs are still predictable whilst a human player is not. The aim of this project is to explore different AI techniques which can possibly make NPCs adapt to the player in the game. One of the AI techniques researched on, the Monte Carlo Tree Search (MCTS), is implemented for the ghosts in the Ms Pac-man game. MCTS is a minimax tree which explores possible paths whilst exploiting the more promising ones. Within 40 milliseconds the MCTS should simulate possible ways for the ghosts to reach Ms Pac-man from their current locations by calculating the reward or value of each path, and finally selecting the highest-rewarded action. The MCTS does not provide the most aggressive ghosts which expert players may gain enough entertainment from. Consequently, a heuristic function is used to provide this level of aggression when Ms Pac-man is playing very well, whilst MCTS is still adopted with varying simulation tree depths based on Ms Pac-man's performance. As a result, the game can still be adaptive by accommodating different skill levels and reacting accordingly to the players.