Multi-Agent Reinforcement Learning: Collaborative Agents in a museum robbery

Abstract

Non-Playable Characters (NPCs) in video games play an important role in the development of an immersive video game. Traditionally NPC behaviours are hard-coded, causing human players to easily identify their weaknesses. Generating NPC behaviours with Reinforcement Learning can introduce real time reactions against human players. In this project, we introduce the early results of a multi-agent team's strategy using Reinforcement Learning techniques in a Non-Zero Sum adversarial asymmetric game. We constructed a complex environment that simulates a museum robbery. The successfully trained team is that of robbers, whose goal is to steal valuables from the museum and leave before being noticed by moving security guards and cameras. The robber team consists of two NPCs with different skills. One is called the Locksmith and is tasked with opening doors and the other the Technician tasked with disabling security cameras. We trained each agent with a different policy while providing them with both individual and group reward signals. The agents learn to cooperate while using their skills for both their own and their team's benefit.