def analyze_position(self, board):
"""Given a game board, analyze the position by evaluating the network on the legal moves
Returns a list of (score, move) tuples, sorted from highest to lowest score, where
the sum of scores is 1.
"""
# Get the encoded boards reachable via legal moves from this position
next_board_tensors = []
next_extra_tensors = []
moves = list(board.legal_moves)
for move in moves:
board.push(move)
nbt, net = Preprocessor.board_to_tensor(board)
board.pop()
next_board_tensors.append(nbt)
next_extra_tensors.append(net)
next_board_tensors = np.array(next_board_tensors)
next_extra_tensors = np.array(next_extra_tensors)
# Evaluate the network on the reachable moves
scores = self.board_score.predict(
[next_board_tensors, next_extra_tensors])[:, 0]
# Convert scores to a probability distribution
scores_dist = np.exp(scores / self.move_temp)
scores_dist /= np.sum(scores_dist)
return sorted(zip(scores_dist, moves), reverse=True)
def get_board_score(self, board):
"""Evaluate the network on the current board and return the score"""
nbt, net = Preprocessor.board_to_tensor(board)
return self.board_score.predict([nbt[np.newaxis], net[np.newaxis]])[0,
0]
