A Reinforcement Learning-Based Classification Symbiont Agent for Dynamic Difficulty Balancing

Abstract

AdaptiveSGA is a mechanism for achieving Adaptive Game AI-based Dynamic Difficulty Balancing in games. AdaptiveSGA is based on the Symbiotic Game Agent model and, therefore, leverages the advantages of biological symbiosis. Within the AdaptiveSGA architecture, the classification symbiont agent is responsible for the dynamic difficulty balancing component. Current work proposes the use of a classification symbiont agent that makes use of reinforcement learning to optimise dynamic difficulty balancing in order to match the opponent's skill. Current work also introduces three different types of decision-making algorithms that can be used by decision-making symbiont agents to display different kinds of behaviour. The ability to reproduce different kinds of NPC behaviour forms the adaptive game AI component of AdaptiveSGA. Experimental results showed that the reinforcement learning-based classification symbiont agent can achieve an even game with opponents and can further help minimise the number of draws.