Improving Adaptive Gameplay in Serious Games Through Interactive Deep Reinforcement Learning

Abstract

Serious games belong to the most important future e-learning trends. Yet, balancing the transfer of knowledge (the serious part) and the entertainment (the playful part) is a challenging task. Whereas a commercial game can always focus on creating more fun for the player, a serious game has to ensure its didactic purpose. One of the major problems in this context is the rigidity of standard script-driven games. The script may guarantee the educational success of the game, but due to its rigid storyline and, consequently, its fixed gameplay, the game may also be boring or overtaxing. Adaptive gameplay seeks to overcome this problem by individualizing the storyline, the difficulty of the game or the amount of context information given to the player. Our idea is to use interactive deep reinforcement learning (iDRL) to maximize the individualization with respect to the context information. Although DRL has been applied successfully to automated gameplay it succeeded mainly at optimization-like tasks in largely short-horizon games. Our approach is to augment DRL with human player and trainer feedback in order to direct the learning process. Our goal could be described as a synergistic combination of human involvement in games and DRL as an emergent cognitive system that helps to adapt the game to players’ preferences and needs. We call this approach interactive deep reinforcement learning for adaptive gameplay.