|
BGonline.org Forums
AI wins final match won series 4-1
Posted By: Taper_Mike In Response To: AI wins final match won series 4-1 (AP)
Date: Thursday, 17 March 2016, at 11:08 a.m.
I have not read any of the published information about AlphaGo. What I gleaned from the interviews of DeepMind team members was that the policy network is like the move filter in a backgammon program. Given the set of all possible moves in a given position, the policy network (which is a neural net) selects those moves that should participate in a tree search. It does this without trying to assess which move is better than another. All it does is decide which moves should be ignored.
The value network (which is a separate neural net) accepts a board position as input, along with the player whose turn it is to move next, and produces a estimate of winning chances for that player.
Is the value network only employed on the bottom leaves of a tree search? That might be so, but I do not know exactly when it is employed.
Mike
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.