Multi-Agent Reinforcement Learning for creating intelligent agents in social networks-oriented role playing games
Francisco Martínez-Gil, Eduard Gil-Magraner · Entertainment Computing · 2025
We present a study for building intelligent Non-Playable Characters (NPCs) that act autonomously as main characters in a hand-to-hand combat-themed Role-playing game (RPG). These types of games are very popular in social networking platforms such as Discord, Amino and WhatsApp. Our work aims for creating intelligent characters that learn individual and collaborative game skills behaving as true opponents for human players. We select three reinforcement learning (RL) algorithms (DQN, PPO and QMIX) to be used with different multi-agent approaches: centralized, totally decentralized and centralized training with decentralized execution (CTDE). First, we assess their performance in NPCs against NPCs fighting with a battery of learning tasks varying the configurations of the fighting groups. Secondly, we suggest a learning strategy that separates the fighting learning in two independent processes for each agent: attack and defense, giving better results than learning a single process for the whole combat. Lastly, we present an analysis of human vs. NPC combats and groups of human vs. groups of NPCs based on the number of victories and a Likert scale survey, concluding that our intelligent agents behave better than an average human player. • Multi-agent reinforcement learning creates smart AI fighters in a Discord game. • DQN, PPO and QMIX algorithms are tested in fighting tasks of incremental complexity. • Our intelligent NPCs win against humans in almost all battle configurations. • 20 players rated learned agent behaviors as suitable via Likert survey