Weakly Supervised Behavioral Modeling for Controllable AI Agents in Video Games
The project aims at developing a new type of Reinforcement Learning algorithm that would allow to retain more control over the artificial agent once its training is completed. This framework would combine modern unsupervised modeling techniques to capture the variability of a set of demonstrations and user-defined programmatic functions that can characterise particularly important factors of variations. Ultimately, the user would be able to specify to the learned agent what type of behavior should be executed at test-time. While this contribution could benefit many AI applications involving sequential decision making, it is particularly interesting for the video game industry as such a tool would allow game designers to make better use of machine learning approaches for in-house testing.