water-vapor / alphazero Goto Github PK

The ground truth of optimizer is a prob map instead of a one hot map. So the self play module need to record the prob map and send to optimizer. A possible way to do this is adding a flag argument to calc_move to let this function return a tuple of (move, prob_map).

BTW, at mcts.py:122, _playout should be self._playout
at mcts.py:145, maybe put pass to eliminate syntax error?

Variable state is not a batched numpy array, it is a GameState object. Please refer to the implementation here, function batch_eval_state