A Computer Game Testbed for Modeling Strategic Decision Making

Matthew Giamporcaro, Boston University

Systems that emulate human decision-making in an interactive environment include computer games and training programs. These systems typically employ such techniques as rule-based algorithms and utility function maximization. In certain applications, these traditional approaches have met with great success, with last year’s chess victory by Deep Blue being perhaps the most famous example.

However, such systems have limited utility, because an experienced human operator will eventually discern the machine’s rules and learn to out-maneuver them. Skilled human opponents typically adapt to one another’s “rules,” as well as to unexpected opportunities and set-backs, and are able continually to maneuver around one another in novel ways. In addition, efforts to create sub-expert computer opponents, such as for training purposes, usually consist of a crude stupefaction of the expert program. This often results in a player that makes mistakes no sane human novice ever would, and which is thus sub-optimal for training.

This project seeks to incorporate adaptive neural network models of human cognitive processing into a commercial game of strategy, in order to help build computer systems that more realistically mimic human players. This approach includes studies on human novices, and also draws from the literature on decision-making, reinforcement learning, and classical game theory. The ultimate goal is to develop a computer program that gradually adapts its playing style in a manner more like that of an improving human novice.

Researchers on this project, in collaboration with a commercial computer game manufacturer, have chosen a popular n-player board game as a testbed. Throughout the course of development, new computer game-players will be designed alongside and tested against the manufacturer’s state-of-the-art rule-based system. The manufacturer has agreed to share proprietary software in exchange for early knowledge of novel cognitive information processing models that may be applied to the company’s products.

Abstract Author(s): Gail A. Carpenter, Matthew W. Giamporcaro