Designing NPC Agents using Value Prediction Methods in Reinforcement Learning

One of the grand challenges in video game development is the design of AI for non-playable characters. The most important requirement in design of AI for these characters is to make them challenging and vulnerable at the same time. In other words, game AI should make the game difficult enough but if this difficulty is too much, the players will not enjoy playing the game. Therefore, most of commercial games use simple techniques such as finite state machines and rule-based systems instead of using modern AI techniques such as machine learning. In this paper, a new method based on value prediction methods in reinforcement learning for designing AI of non-playable characters is proposed. The most important advantages of this method are that it is a non-deterministic method and its intelligence level can be easily tuned by the developer. Furthermore, this method is domain-independent and thus, it can be easily used in commercial games. For the sake of evaluation, result of using this method in a 3D soccer free kick simulation game is reported.

کلیدواژه ها:

Non-playable characters ، Game AI ، Reinforcement learning ، Value Prediction Methods