Integrating genetic algorithm and fuzzy integral for evaluating game units combination in RTS game

82 leaves : ill. ; 30 cm.PolyU Library Call No.: [THS] LG51 .H577M COMP 2012 LiYReal time strategy (RTS) games provide challenging test beds for developing human-level artificial intelligence (AI), especially real-time planning and decision making under uncertainty. For the game players, with the limited game play resources, there are often complex interactions among game units and features for him to control in order to achieve the game objectives. In RTS games, the large decision spaces with multiple possible strategy treads lead to complicated calculation. Furthermore, a good game needs to develop in-game AI to control the non-playable characters (NPC) to play with the players, and it is often difficult to develop such good AI. This research focuses on the automatic detection of players' game strategies in data. By detecting and generalizing these game strategy patterns in the available data, a system can be developed to evaluate whether a strategy is appropriate or not. This system can also make predictions about how to further interact with the players so as to keep the game continuously being attractive to them.Unlike many other traditional pattern recognition problems in which the goal is to understand the relations and regularities in some source data for the purpose of predicting the class label of new data coming from the same source, computer game data, usually in the forms of game logs, records the interactions that have been carried out between the player and the game system. In RTS games, these interactions are particularly interesting. They describe how the player plans, develops and implements strategies to deal with his opponents (e.g., the game system or other online players). The RTS game data consists of (i) various players' actions such as attacks and retreats, building creation sequences, (ii) number of combat units and their types and conditions, (iii) information about the environment, and the like. Moreover, substantial feature interactions are being found among these game units, the traditional machine learning algorithms based on weighted sum on features importance are no longer satisfactory. Advanced measuring techniques are needed. Finally, game data also doesn't possess the concept of a concrete "class label", thus a new way of defining accuracy has to be formulated. In this thesis, a machine learning approach that integrates genetic algorithm, fuzzy measures and integrals is developed to extract case strategies for evaluation. Fuzzy measure is used to incorporate the complex interactions among different unit types in fuzzy integral. These fuzzy measures are found using genetic algorithm. A new fuzzy integral with respect to the RTS game rules has been developed. The new fuzzy integral relaxes the monotonicity requirement and can deal with the production order of game units. Real-life game data of Warcraft III are used for testing and evaluating this approach. Compared with the traditional Choquet Integral, our approach obtained a promising result from the experimental testing.Department of ComputingM.Phil., Dept. of Computing, The Hong Kong Polytechnic University, 201