於模擬賽車遊戲學習超車與阻擋行為之研究

Abstract

這篇論文的內容是在研究TORCS平台上，以學習演算法訓練出來的賽車AI，在不同阻擋者以及不同軌道條件下的表現。論文一開始會先從為何我們想要研究更具侵略性的AI駕駛行為開始講起，接著介紹TORCS這個開放程式碼的賽車遊戲平台架構以及在各個研討會或期刊上與TORCS相關的研究。因為TORCS所內建的賽車AI其駕駛行為皆不具侵略性，在超車時如果剩餘的軌道寬度沒有達到一定的安全寬度，則會取消超車行為，且內建AI並不存在阻擋行為，若後車速度較快，前車會將路徑讓給後車。所以我們利用學習演算法中的Q-learning algorithm來發展更具侵略性的超車與阻擋行為，並統計在不同情形下賽車的數據來評比表現。

In this thesis, we research the learning of overtaking and blocking behaviors of AI (non-human) drivers on TORCS, a simulated car racing platform. At first, we introduce why we focus on the more aggressive driving strategy. Secondly, we introduce the TORCS racing game architecture and some researches in TORCS or AI-related conference. Because of built-in AI didn’t have aggressive driving skill, when performing overtaking, if remain track width has not reach to a safely width, the built-in AI will cancel the overtaking behavior. Additionally, the built-in AI didn’t have blocking behaviors. If the opponent’s speed was higher than theirs, it will let the opponent pass. In order to improve built-in AI’s driving strategy, we use Q-learning algorithm for learning behaviors and analysis their overtaking and blocking performance in different conditions.