EVOR: An online evolutionary algorithm for car racing games

Abstract

In this paper, we present evolutionary racer (EVOR) that is a simulated car dynamically controlled by an online evolutionary algorithm (EA). The key distinction between EVOR and earlier car racing methods is that it considers car racing as a dynamic optimization problem which is addressed by an evolutionary algorithm. Our approach calculates a car trajectory based on a controller decision and adjusts this decision according to the suitability of its resultant trajectory with the current track status. Furthermore, it allows to integrate features such as opponent handling implicitly. Our experimental results show that EVOR outperforms current best AI controllers on a wide range of tracks.