<![CDATA[The gaming industry has grown rapidly, and one of the key elements in a game is the non-playable character (NPC). Easily predictable NPC behavior often reduces player engagement and satisfaction. Static or unresponsive NPCs tend to create monotonous and less challenging gameplay experiences, ultimately lowering game quality and player interest. This study applies Deep Reinforcement Learning (DRL) using Unity ML-Agents to train agents in a Dodgeball game, enabling them to make adaptive decisions through self-play. A reward system was designed to provide positive feedback for strategic actions, such as picking up and throwing the ball, and penalties for mistakes, such as hitting walls or being hit by the ball. The training results showed a gradual improvement in agent performance, reflected in the increasing and stable cumulative rewards and ELO scores at the end of training. In performance testing, the DRL agent achieved a 66% win rate against the rule-based agent over 50 matches. A user preference test also revealed that 80% of players preferred competing against the DRL agent, with 60% of them considering it more challenging than the rule-based one. These results demonstrate that the DRL agent not only outperforms the rule-based agent but also provides a more dynamic and realistic gameplay experience.]]>
Copyrights © 2025
