الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 72 |  |  |
| | | from | 72 | | |
Abstract
Real-time strategy games are strategic war games where two or more players operate on a virtual battlefield, controlling resources, buildings, units and technologies to achieve victory by destroying others. Achieving victory depends on destroying the opponent, destroying the opponent depends on selecting a suitable plan or strategy (set of actions), selecting a suitable plan depends on building an imagination (building a model) of the opponent to know how to deal with. This imagination is the opponent model, the stronger the opponent modeling process is, the more accurate the selected suitable plan is and consequently the higher probability achieving the victory is. Creating a human-like computer player in real-time strategy games requires huge number of opponent models, these models must be preprocessed to either focus on accuracy or performance according to our needs. In order to preprocess these models accurately, we need to detect their type. Opponent models’ type can be complex or simple. Complex opponent models are low variance models whose differences in features’ values are low, so in order to accurately separate between these models, we need to preprocess them by increasing their dimensions. Simple opponent models are high variance models whose differences in features’ values are high, so in order to separate between these models in a reasonable time, we need to preprocess them to decrease their dimensions, if possible, without accuracy or data loss. Finally, the new observed model is classified with the preprocessed training opponent models using the corresponding classifier according to the detected type of the preprocessed training opponent models.
The problem is that most recent researches propose methodologies that either focus on time performance or accuracy, another problem is that many classification processes are game-specific. In this thesis, a new methodology is proposed that focuses on both time performance and accuracy in building and classifying the opponent model. Our new methodology can also classify the observed opponent model in a way that is not game-specific. The methodology mainly includes two paths, only one of them is executed per opponent models trained, which means that different types of real-time strategy games and different races may execute different paths of the two paths of our methodology. Results show the error rate, the enhancements performed by the classification processes and the time performance.
Keywords: Real-time strategy games, opponent modeling, classification, clustering, opponent modeling, adaboost, multiboost, svm, pca, neural networks, rough sets, rts.