Igor Butinar (2013) Parallelization of Monte Carlo Tree Search. EngD thesis.
Abstract
The main objective of this graduation thesis is the parallelization of the program, the core of which is the UCT algorithm and is designed for playing various games. The program is written in the C++ programming language. I got the original program from the Laboratory for Adaptive Systems and Parallel Processing (LASPP), where my mentor on this thesis is a member. I used the MPI interface to parallelize the program. Because of completeness I described the essential features, components and concepts that are used by MPI (e.g. sending and receiving messages, collective communication, communicator, group, process, data type, one-sided communication, etc.). I also described most of MPI functions and provided some examples of their use in the C programming language. I described the MCTS family of algorithms, of which the UCT algorithm is a part of, together with some improvements and extensions. In the end I performed measurements, where I observed and analyzed the performance improvements and identified which parallel implementation yields better results regarding games with the processing time as the key metric.
Actions (login required)