The idea that top-level programs use UCT is just folklore at this point. I know that at least MoGo uses a greedy version of Monte-Carlo Tree Search enhanced wit

The idea that top-level programs use UCT is just folklore at this point. I know that at least MoGo uses a greedy version of Monte-Carlo Tree Search enhanced with RAVE (Rapid Action Value Estimation), a way to share information between branches of the search tree, a pattern database which runs an evaluation function on small patches of the board, and a set of simple heuristics. The consensus is that UCT is bad because it tries to minimize cumulative regret (the average performance over all simulations) at each node, rather than the strength of the chosen move, and that the various heuristic strategies used by strong engines introduce enough noise to make greedy playouts the better option. If you want to learn more, you might want to check out these two papers: http://www0.cs.ucl.ac.uk/staff/D.Silver/web/Applications_files/mcrave.pdf http://www0.cs.ucl.ac.uk/staff/D.Silver/web/Applications_fil... http://arxiv.org/pdf/1206.3382.pdf http://arxiv.org/pdf/1206.3382.pdf