Δενδρική αναζήτηση Monte Carlo στο παιχνίδι στρατηγικής "Άποικοι του Κατάν" (EL)

Monte Carlo tree search in the "Settlers of Catan" strategy game (EN)

Καραμαλεγκος Εμμανουηλ (EL)

Karamalegos Emmanouil (EN)

Δεληγιαννακης Αντωνιος (EL)

Λαγουδακης Μιχαηλ (EL)

Χαλκιαδακης Γεωργιος (EL)

Πολυτεχνείο Κρήτης (EL)

Chalkiadakis Georgios (EN)

Lagoudakis Michael (EN)

Technical University of Crete (EN)

Deligiannakis Antonios (EN)

Classic approaches to game AI require either a high quality of domain knowledge, or a long time to generate effective AI behavior. Monte Carlo Tree Search (MCTS) is a search method that combines the precision of tree search with the generality of random sampling. The family of MCTS algorithms has achieved promising results with perfect-information games such as Go. In our work, we apply Monte-Carlo Tree Search to the non-deterministic game "Settlers of Catan", a multi-player board-turnedweb- based game that necessitates strategic planning and negotiation skills. We implemented an agent which takes into consideration all the aspects of the game for the first time, using no domain knowledge. In our work, we are experimenting using a reinforcement learning method Value of Perfect Information (VPI) and two bandit methods, namely, the Upper Coefficient Bound and Bayesian Upper Coefficient Bound methods. Such methods attempt to strike a balance between exploitation and exploration when creating of the search tree. For first time, we implemented an agent that takes into consideration the complete rules-set of the game and makes it possible to negotiate trading between all players. Furthermore, we included in our agent an alternative initial placement found in the literature, which is based on the analysis of human behavior in Settlers of Catan games. In our experiments we compare the performance of our methods against each other and against appropriate benchmarks (e.g., JSettlers agents), and examine the effect that the number of simulations and the simulation depth has on the algorithms’ performance. Our results suggests that VPI scores significantly better than bandit based methods, even if these employ a much higher number of simulations. In addition to this, the simulation depth of the algorithm needs to be calculated so the method will neither get lost in deep simulations of improbable scenarios neither end up shortly without given a proper estimation of the upcoming moves. Finally, our results indicate that our agent performance is improved when the alternative, human behavior-based, initial placement method. (EN)

bachelorThesis

Artificial intelligence (EN)

Πολυτεχνείο Κρήτης

Ιδρυματικό Αποθετήριο Πολυτεχνείου Κρήτης

School of Electrical and Computer Engineering - Diploma Works

Intelligent Systems Laboratory - Diploma Works

Αγγλική γλώσσα

2016

http://purl.tuc.gr/dl/dias/1CCC4417-B1AA-4C2E-B33B-0F6BEFDED36Ehttps://drive.google.com/drive/folders/0B9pKS-Z7bpHxWEotSUxFbGhFUWM

Πολυτεχνείο Κρήτης::Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών (EL)

Technical University of Crete::School of Electrical and Computer Engineering (EN)