Research On Risk And Opponent Modeling In Imperfect Information Game

Artificial intelligence (AI) is an important branch of computer science. It targets to do study on how to make the computer to master works that rely on human intelligence. As one of the research areas of artificial intelligence, computer game is a crucial approach of measuring AI system, which brings a lot of important methods and theories to artificial ntelligence, and provides a wide range of social and academic influence.Games can be classified according to whether or not they are perfect or imperfect nformation. In an imperfect information game, players have non-singleton information ;et which means they have only partial knowledge about the state of the game. These nake the studies on imperfect information games more complex, competitive and mean-ngful. Compared with those of perfect information, the players of imperfect information vill surely suffer risk lost in their strategy decision process. And also, strategy deci-sion follows different Nash Equilibrium because of the individuation of players’ actions. Therefore, it is of great value to study these theses, which are an important and difficult )art in computer game field.The research of this paper will focus on the research of imperfect information game ystems’risk and opponent modeling. The main points are searching method about great cale game tree; risk estimation and prevention and opponent modeling.Imperfect information conditions led to the great scale of game tree. In this area, vlonte-Carlo Tree Search (MCTS) approach has been proved to be efficient and wildly ipplied to solve of the great scale game tree searching problems. UCT algorithm provides ecision policies in the process of game tree branch selection. In this paper, different policies of UCT algorithms are realized and compared in our experiments. The regulars imong policies and characters specific problems are concluded.In imperfect information conditions, there are always deviations between player’s xact payoff and his expectation. These deviations come from the inaccurate prediction environment and opponent information. In this thesis, based on the analysis of perfect ind imperfect factors in game problems, the concept of risk lost is presented and the estimation method is provided. Then, a modified UCT policy, UCT-Risk is raised and confirmed its advantage in problems with high risk factors. Based on these, a systematical model of strategy selection of risk dominance is proved at last.The imperfect and asymmetrical information leads players of different strategy de-cision equilibrium. Building opponent models, analyzing the individuation and clustering of opponents and building more effective strategy decision model become important areas in recent years. Thus, opponent model theory and approaches are deeply studied in this hesis. The method of building opponent models in board game is provided systemati-cally.Opponent clustering problem is further discussed in this thesis. Classic machine earning methods always require high quality and sufficiency of history data. Thus, K-neans clustering method is applied and modified in this thesis. KL Divergence is used as he criterion of data clustering. This thesis studies the ability of history analysis to reduce iability of history data while maintaining the accuracy.At last, the imperfect information game system evaluation method and data collec-ion are discussed. A computer game experiment platform is built which can support great cale human-machine competition. Based on the collected data, Machine Learning meth-)ds are applied to train the estimation function of game system, which further improve game system’s performance.