Application Of Ai In The Game: Augumenting Game Player's Experience Using Real-Time Player Modeling And Opponent Adaptation

The motivation of research on AI (Artificial Intelligence) in Video Games is different from that of in Classic Game. For example, the research on the classic game AI of Go is to produce the most challengeable opponent intelligence; while, the research on Video Game AI is to generate the both challenging and satisfying game intelligence. The "game" in the present research refers to video games. Most existing game AI is implemented by FSM (Finite State Machine) which has drawbacks in the three respects:highly requirement of designer's domain knowledge and participation; no existence of meta-programming; no planning and looking forward.In the thesis, the author proposed the approach of CI (Computational Intelligence) to generate both challengeable and satisfactory game opponent, under the proposed framework of "Player Modeling and Opponent Adaptation". At the present research, CI refers to MCTS (Monte Carlo Tree Search) and UCT (Upper Confidence Bound for Trees); two prey/predator genre games of Dead-End and Pac-Man are used as test-beds to validate the proposed theory.Contributions of this thesis includes:"Player Modeling and Opponent Adaptation" as a framework of game AI; strategy-based player modeling; opponent's challengeable adaptation; and opponent's satisfactory adaptation.On the research of game AI framework, the author proposed "Player Modeling and Opponent Adaptation" as a framework, which is proved feasible for prey and predator genre games AI of Dead-End and Pac-Man, and is possible to be used for future game AI developments.On the research of how to model player's strategy, the author first proposed the approach of using supervised learning for modeling player's strategy during the gameplay (classification or pattern recognition). The procedure for it include:attribute collection, sample data collection, noisy data processing, attribute subset selection, choice of classification algorithm, and training and evaluating the classifier. In addition, the author proposed the approach of using unsupervised learning for clustering player's strategies, classification or recognition, and evaluation. The procedure for it include:attribute collection, sample data collection, choice of clustering algorithms, data grouping, and choice of a Cluster algorithm with cross validation test.On the research of how to generate adaptive opponent which are challengeable, the author proposed straight-CI to control NPC (non-player character) in order to generate challengeable intelligent opponent. The advantage of CI is that it rarely requires human participation and it depends little on domain knowledge, so CI could also be used for AGD (Automatic Game Design/Development). Disadvantage of CI is that it needs to be online, resource consuming, so it applies only to standalone PC games. In addition, the author proposed knowledge-based-CI to control NPC in order to generate challengeable intelligent opponent. This approach uses straight-CI to control NPC in order to generate data, which is used for training the ANN (artificial neural network). The trained ANN later can be used for control of NPC. It has the advantage of trained offline, knowledge-based, high performance and resource efficient, so knowledge-based CI is more feasible for Multi-player Online Games (MOGs). However this approach is required to be correlated with Strategy-Based Player Modeling. These two aforementioned CIs both outperform FSM for implementing game AI in the above mentioned three aspects.On the research of how to generate adaptive opponent which are satisfactory, we try to optimize the player's experience through the creation of an even game AI. To better satisfy the player, we proposed two CI-based DDAs (Dynamic Difficulty Adjustment):"DDA by time-constrained-CI" and "DDA by knowledge-based-time-constrained-CI". As the latter is based on knowledge and more computational resource efficient than the former, so it is more applicable for multi-player online games, while the former is only applicable for standalone PC game. In addition, the latter are required to accomplish under the framework of "Player Modeling and Opponent Adaptation". Both of the DDAs by CI outperform the existing DDAs. As at the present, the existing DDAs are accomplished by adjusting the speed and number of NPCs, this can frustrate the players; while the DDA by CI is accomplished through the adjustment of game intelligence (as we found that with the increment of CI simulation time, the performance of the game AI improves), which is easy to optimize the player. These two DDAs are based on the result of aforementioned research on "to generate adaptive opponent which are challengeable".