Computer Agent for Board Game Playing

U ovom radu proučeni su različiti pristupi oblikovanja računalnog agenta za automatsko igranje odabrane igre na ploči. Opisan je postupak efikasne implementacije igre kao i razvoj grafičkog korisničkog sučelja za evaluaciju agenta u igri protiv ljudskih igrača. Implementirani su algoritmi iz dvije različite familije: tradicionalni algoritmi koji se temelje na klasičnom minimaks pretraživanju stabla te noviji koji koriste Monte Carlo simulacije i duboko učenje. Razvijeni agenti su iterativno poboljšavani raznim optimizacijskim postupcima sve dok nije postignuta željena razina kvalitete igre. Posebno su naglašene mogućnosti, ograničenja i problemi pojedinih algoritama. Na koncu, nakon evaluacije svih razmotrenih algoritama, odabran je najbolji te je predložena njegova konačna natjecateljska implementacija.This paper considers different approaches of designing computer agents for automatic board game playing. It describes the process of efficient game engine implementation as well as graphical user interface for evaluating agent game play against human players. Algorithms from two different families were implemented: traditional algorithms which are based on classical minimax tree search and new ones which utilize Monte Carlo simulations and deep learning. Developed agents had been iteratively enhanced with various optimization techniques until satisfactory quality of game play was reached. Special focus was put on possibilities, limitations and problems of chosen algorithms. At last, after evaluation of all considered algorithms, the best one was chosen and its final competitive implementation was suggested