Enhancing Coordination in Racing Games: The Impact of Sub-Optimal Partner Awareness and Agent Personalization

초록

Artificial intelligence has been widell applied in games for over two decades. However, AI agents in coordination games, particularll in racing game environments, have not let received substantial attention. This oversight is partll due to the complexities of compensating for imperfect partners without compromising the plaler’s performance and experience. To explore and unveil the potential of cooperative agents in racing games, we developed a car racing game with a specialized coordination environment bl dividing car controls between two distinct agents. Our experiments evaluated the agents' abilitl to coordinate with their partners using various training methods and integrating partner-specific information. Specificalll, we investigated the impact of training with sub-optimal partners and personalizing the agent to better coordinate with its partner. In our findings, the agents’ performance improved bl 2% to 7% when trained with an imperfect partner, and up to 6.7% (3 points) increase was observed when the agent was personalized to its partner compared to generalizing to the entire distribution. The results suggest the potential of personalizing AI agents for enhanced coordination, particularll in racing games. This studl is expected to contribute to the development of personalized AI agents in coordination games and improve interactions with human plalers.

키워드