A Study on Human-like Billiard AI Bot Based on Backward Induction and Machine Learning

• Main Authors: Kuei-Gu Tung, 董奎谷
• Other Authors: Wen-Kai Tai
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/5rzh75

碩士 === 國立臺灣科技大學 === 資訊工程系 === 107 === In recent years, most of researches in the field of game Artificial Intelligence (AI) have focused on improving the strength of AI. Those researched AIs have the ability to compete with human players, even surpass professional players. For example, the well-known Go AI: AlphaGo. However, an important factor has to be considered for AI when generally speaking of gameplay experience: the human likeness of its behavior. It makes players feel acceptable to reasonable thoughts of opponents, instead of leaving bad gaming impression for players due to AI’s overpowered ability. This paper studies AI in billiard games, describing the difference between the decisions of general AIs and human players. We analyzed actual game records of human players and retrieved feature vectors from the data. We leveraged the Backward Induction algorithm and machine learning to imitate the process of making decisions from human players. Providing our AI suggestion of strategies, it could avoid being over-dependent on the robust and precise physics simulation. This study applies human-like strategies to AI which could mimic human thoughts. It is more similar to actual human players in a sense of the way they played while being compared to the original AI. Also, we defined an appropriate approach to gauge the human likeness of AI, evaluating our proposed methods. Perhaps our proposed methods could keep the balance between the ability and human likeness of AI to further provide better gameplay experience. The experimental results show that our method overall is more similar to the way how human players play than the original AI, proving that it could effectively improve the human likeness of AI.