def _choose(self, state, actions):
if self.is_train_mode and random.random() < self.exploit_rate:
next_pos = random.choice(actions)
if self.debug: print("SELECT", actions, "RANDOM", next_pos)
return next_pos
found_p = -1.0
found_c = []
ob = OptimalBoard(state)
_id = ob.board_id
if self.debug: print("FROM", _id)
scores = self.p_table.lookup(_id)
converted_actions = ob.convert_action_to_optimal(actions)
for action in converted_actions:
p = scores[action]
if self.debug: print("ACTION", ob.convert_action_to_original(action), p)
if p > found_p:
found_p = p
found_c = [ob.convert_action_to_original(action)]
elif p == found_p:
found_c.append(ob.convert_action_to_original(action))
next_pos = random.choice(found_c)
if self.debug: print("SELECT", found_c, found_p, next_pos)
return next_pos
def _choose(self, state, available_actions):
if self.is_train_mode and random.random() < self.exploit_rate:
return random.choice(available_actions)
ob = OptimalBoard(state)
converted_actions = ob.convert_action_to_optimal(available_actions)
action = self.q.rargmax(ob.board_id, converted_actions)
return ob.convert_action_to_original(action)
def _choose(self, state, available_actions):
optimal_board = OptimalBoard(state)
converted_actions = optimal_board.convert_action_to_optimal(
available_actions)
converted_state = self.convert_state(optimal_board.optimal_board)
###
if self.is_train_mode:
if random.random() < self.egreedy:
action = random.choice(converted_actions)
else:
action = self.network.predict_one(converted_state)
else:
action = self.network.predict_one(converted_state)
if action not in converted_actions:
# 여기에 뭐를 학습으로 넣을 지 고민
# 아니면, predict_one에서 필터를 넣을 지 고민
self.network.add_train_set(converted_state, action, -1,
self.convert_state([-1] * 9), True)
action = random.choice(converted_actions)
original_action = optimal_board.convert_action_to_original(action)
return original_action