Learning Non-Differentiable Graphs of Utility AI

Abstract

Utility AI is an approach to modelling AI players in computer games. Its structure is a graph that computes the utility values of possible actions and chooses the one with the highest value. Currently, such graphs are created by experts manually. This paper presents the first attempts to create them automatically – through learning from data. The problem is similar to training neural networks except that the utility graphs are non-differentiable and contain various types of nodes (more complex than neurons). We present the most promising methods, preliminary experiments and results.