Interpretable Utility-based Models Applied to the FightingICE Platform

One task of game designers is to give NPCs fun behaviors, where “fun” can have many different manifestations. Several classic methods exist in Game AI to model NPCs' behaviors; one of them is utility-based AI. Utility functions constitute a powerful tool to define behaviors but can be tedious and time-consuming to make and tune correctly until the desired behavior is achieved. Here, we propose a method to learn utility functions from data collected after some human-played games, to recreate a target behavior. Utility functions are modeled using Interpretable Compositional Networks, allowing us to get interpretable results, unlike regular neural networks. We show our method can handle noisy data and learn utility functions able to credibly reproduce different target behaviors, with a median accuracy from 64.5% to 83.7%, using the FightingICE platform, an environment for AI agent competitions. We believe our method can be useful to game designers to quickly prototype NPCs' behaviors, and even to define their final utility functions.