How to advance general game playing artificial intelligence by player modelling
This addresses the problem of creating more human-like and generalizable AI for game playing, though it is incremental as it builds on existing player modelling methods.
The paper argues that current deep learning approaches to general game playing AI have limitations in data dependency, problem construction, and lack of human-level complexity, and proposes that a generalised player model based on category theory and psychology is needed to advance the field.
General game playing artificial intelligence has recently seen important advances due to the various techniques known as 'deep learning'. However the advances conceal equally important limitations in their reliance on: massive data sets; fortuitously constructed problems; and absence of any human-level complexity, including other human opponents. On the other hand, deep learning systems which do beat human champions, such as in Go, do not generalise well. The power of deep learning simultaneously exposes its weakness. Given that deep learning is mostly clever reconfigurations of well-established methods, moving beyond the state of art calls for forward-thinking visionary solutions, not just more of the same. I present the argument that general game playing artificial intelligence will require a generalised player model. This is because games are inherently human artefacts which therefore, as a class of problems, contain cases which require a human-style problem solving approach. I relate this argument to the performance of state of art general game playing agents. I then describe a concept for a formal category theoretic basis to a generalised player model. This formal model approach integrates my existing 'Behavlets' method for psychologically-derived player modelling: Cowley, B., Charles, D. (2016). Behavlets: a Method for Practical Player Modelling using Psychology-Based Player Traits and Domain Specific Features. User Modeling and User-Adapted Interaction, 26(2), 257-306.