Population-Based Evaluation for Dynamic Difficulty Adjustment in Repeated Rock-Paper-Scissors

Dynamic Difficulty Adjustment (DDA) aims to personalize game challenge by adapting AI behavior to player skill, supporting engagement and sustained motivation. In this work, we present a population-based evaluation framework for DDA in Repeated Rock-Paper-Scissors. Agents are trained against a diverse population of rule-based bots to achieve a prescribed win-loss-draw distribution. Experimental results demonstrate that population-based RL enables agents to closely match target outcome distributions and generalize well to novel opponents, illustrating the promise of our approach for scalable, robust DDA.