Deep Reinforcement Learning and Influenced Games

In game design, creators strive to guide players toward specific strategies through the game's mechanics. This assumes a level of predictability in player behavior, though in reality, players often deviate from expected rational strategies. We provide an ensemble mechanism to influence behavior in arbitrary normal-form games, even when players are not rational. Our algorithm adjusts the game's reward structure to encourage gameplay that aligns with the designer's intentions. At the algorithm's core is a deep reinforcement learning technique that learns to model players' real-world behavior. This mechanism is versatile and applicable beyond game design; it can be employed in various fields where one can frame problems as classification tasks. Incorporating actual gameplay data into the training process allows our algorithm to acquire a practical understanding of player decisions. The efficacy of our mechanism is evaluated by testing the mechanism's performance against a panel of classifiers, which includes a support vector machine, random forest, and multi-layer perceptron.