Monte Carlo tree search in a distributed environment

Jure Grabnar (2015) Monte Carlo tree search in a distributed environment. EngD thesis.

Preview

PDF Download (596Kb)

Abstract

Monte Carlo Tree Search algorithm (MCTS) is a computationally expensive algorithm. The time needed for computation correlates with the quality of the results. The purpose of this work is to parallelize MCTS method. With parallelization we gain an ability to increase the number of simulated games per turn and other parameters and still be able to receive results in sufficient time. Quality of results has been improved significantly. Parallel algorithm was written in MPI library which enables the program to run on multiple computers. Algorithm was evaluated on different problem sizes. With big enough problem, the speedup was approximately linear. Algorithm was run on a grid network which is administered by Slovenian Initiative for National Grid (SLING). As a part of this work, instructions for usage of grid network were created.

Actions (login required)