def build_policy(self, action, flip):
labels_n = len(ActionLabelsRed)
move_lookup = {move: i for move, i in zip(ActionLabelsRed, range(labels_n))}
policy = np.zeros(labels_n)
policy[move_lookup[action]] = 1
if flip:
policy = flip_policy(policy)
return policy
def action(self, env: CChessEnv) -> str:
value = self.search_moves(env) # MCTS search
policy = self.calc_policy(
env) # policy will not be flipped in `calc_policy`
if not env.red_to_move:
pol = flip_policy(policy)
else:
pol = policy
my_action = int(
np.random.choice(range(self.labels_n),
p=self.apply_temperature(pol, env.num_halfmoves)))
# my_action = np.argmax(self.apply_temperature(pol, env.num_halfmoves))
# no resign
self.moves.append([env.observation, list(policy)
]) # do not need flip anymore when training
return self.labels[my_action]
