Application of machine learning algorithms to character interaction in single-player role-playing games

Abstract

Character interaction in single-player role-playing games refers to the role interactions between the player character and the non-player characters (NPCs) in the game including dialogue, social interaction, cooperation, and competition, which are designed and implemented to achieve certain game goals and effects. This paper summarizes and compares machine learning algorithms applied to game development using a paper review and theoretical analysis, aiming to provide new solutions for optimizing and improving algorithms commonly used in game design. The ADEM algorithm can automate the evaluation of semantic relevance and logical coherence of character dialogues, but it is more costly in the pre-training phase and more difficult to perform domain adaptation for specific game content. The BERT algorithm can improve the speed of character dialogue generation through the training of unlabeled text, and then improve the flexibility of the dialogue through a small amount of labeled text for domain adaptation and fine-tuning, but it lacks in the generation of smooth and complex text and other aspects. The GAN algorithm can learn from a single or a small number of action sequences to achieve the transformation and mixing of character actions, thus generating multiple action gestures. However, its training process may suffer from problems such as pattern collapse leading to degradation in the quality of the generation, and it is also difficult to set a uniform objective criterion to determine the loss function.