Building a Framework for AI agent-based modeling of player performance style during a steering task

Abstract

This thesis introduces the development of a framework for artificial intelligence agent-based modeling of player performance style during a steering task. The framework originates from a game called CogCarSim, developed to study flow — a psychological state in which a person is fully immersed in an ongoing task. The CogCarSim is a "custom-made high speed steering game", but it neither has the ability to manipulate players’ experiences nor modeling players capability. The current thesis presents steps in developing the original game into a modular framework that supports adding functionalities to control players’ experiences (custom maps, speed gates) and modeling players (state-space representations, search agents). The thesis first covers background knowledge on flow, play, the relationship between flow and play, and player modeling. To achieve player modeling competence, the framework requires an artificial intelligence that can be tailored to represent different playing styles, reflecting human biases and preferences. For that purpose, we choose Monte Carlo Tree Search (MCTS) over Reinforcement Learning or other search algorithms (such as Breadth First Search, Minimax, or Depth First Search). The overall aim of the thesis is to introduce a new framework to study the flow phenomenon and have the framework model players successfully. As a result, the work produced a framework with a number of core components required towards this aim. Yet, the implemented search agent did not produce expected results due to challenges in hyperparameter tuning, challenges in mapping discrete game-space coordinates to a representative decision making and action space, and potential issues in the interaction between the MCTS and support functions, and these issues would require further investigation. However, the framework laid a strong foundation for future development and study.