def get_action(self, board, temp=1e-3):
#sensible_moves = board.availables
# the pi vector returned by MCTS as in the alphaGo Zero paper
if self._is_selfplay:
temp = 1.5
move_probs = np.zeros(15 * 15)
acts, probs = self.mcts.get_move_probs(board, temp)
if acts is None: #ai认输
return None, None
move_probs[list(acts)] = probs
best_chance = np.max(move_probs)
best_move = np.where(move_probs == best_chance)[0][0]
if self._is_selfplay:
move = np.random.choice(
acts,
p=probs
#p=0.9*probs + 0.1*np.random.dirichlet(0.3*np.ones(len(probs)))
)
#debug
print("choose ", RenjuBoard.number2pos(move), "by prob ",
move_probs[move])
print("best move is ", RenjuBoard.number2pos(best_move),
best_chance)
# update the root node and reuse the search tree
else:
# with the default temp=1e-3, it is almost equivalent
# to choosing the move with the highest prob
#move = np.random.choice(acts, p=probs)
move = best_move
# reset the root node
#self.mcts.update_with_move(-1)
self.mcts.update_with_move(board, move)
return move, move_probs
def _debug(self):
if self.debug_mode:
for act, _sub_node in self._root._children.items():
if _sub_node._n_visits > 0:
print(RenjuBoard.number2pos(act), "\tsel ",
_sub_node.get_value(self._c_puct), "\tv ",
_sub_node._n_visits, "\tQ ", _sub_node._Q, "\tp ",
_sub_node._P)