April 2018
Intermediate to advanced
334 pages
10h 18m
English
Input of the board representation is received, which is a 19 x 19 x 17 tensor. It is passed through a residual convolution network then fully connected layers finally output a policy vector and a value representation. Initially, the policy vector will contain random values since the networks start with random weights initially. Post obtaining the policy vector for all possible moves for the given state, it selects a set of possible moves having very high probabilities, assuming that the moves having the high probabilities are also potentially strong moves:
