Game kernel design using self organizing map

Abstract

In this project, a game kernel design using artificial neural network (ANN) was constructed. The neural network used in this project was Kohonen self-organizing map (SOM). This network was used as a classifier to classify different input patterns that were inputted to train this network. The input patterns were then clustered into different categories. The input patterns were derived from the game data gathered in game playing by a good player. After the training of the network, the network learnt some strategies of the player. A C++ written tank battle game was used for the implementation of this game kernel. The game was a fight of two tanks in a rectangular area. The actions these tanks could take are attack, dodge, follow, evade, go to center etc. The resulting artificial intelligence (AI) of this game was tested in a fight with one tank using this AI and another tank with random actions. Our AI was able to defeat the random AI easily. Also, the resulting AI was evaluated by inviting volunteers to play the game and give comments. 75% of volunteers agreed that our AI was much interesting than the randomly generated AI. Finally, our AI was tested with its player modelling ability by using a player modelling training set and a program generated training set to calculate their fitness values with respect to a test set which was a set of training data gathered in game play. Our AI was found to be able to model 44% of player strategies using player modeling training set while the AI using program generated training set modeled only 16% of the player’s strategies.