The impact of transfer learning on the learning efficiency and performance of an AI agent in a game environment

Sammanfattning

In the evolving landscape of artificial intelligence within digital gaming, the creation of dynamic and responsive environments remains a paramount challenge. However, the advent of machine learning (ML) and, more specifically, transfer learning, presents a novel paradigm in the development of game AI. Transfer learning, a facet of ML, involves the adaptation of models trained on one task to perform another related task and unlike conventional techniques, which are often limited by their predefined rules, transfer learning enables AI agents to assimilate and evolve, using knowledge from previous training. The potential of transfer learning in games extends beyond mere behavioural complexity; it signifies a leap towards truly immersive and interactive gaming worlds where AI entities can learn, evolve, and respond in ways previously confined to the realm of human players. In this thesis, we present an experiment where we explored how instant and delayed rewards affect AI agents’ learning efficiency and performance when transferred to new gaming environments. For the experiment, our research utilised a proprietary game engine and two procedurally generated top-down 2D games, each with unique challenges and objectives. Two models were trained on the first game with either instant or delayed rewards respectively, transfer learning was then applied by adapting the models to the second game. Our results showed that although the pre-transfer model with instant rewards performed higher than its counterpart, the transferred models differed. Specifically, the transfer model based on delayed rewards yielded the strongest performance, demonstrating that the long-term advantages of the general model surpass the short-term benefits offered by the environment-specific model.