Computer Go is one of the most challenging field in Artificial Intelligence Game. In this area the use of Monte Carlo Tree Search has emerged as a very attractive research direction, where recent advances have been achieved and permitted to significantly increase programs efficiency. These enhancements result from combining tree search used to identify best next moves, and a Monte Carlo process to estimate and gradually refine a position accuracy (estimation function). The more the estimation process is accurate, the better the Go program performs. In this paper, we propose a new approach to extract knowledge from the Go tree search which allows to increase the evaluation function accuracy (BHRF: Background History Reply Forest). The experiments results provided by this new approach are very promising.