def respond(self, env):
mask = get_mask(to_char(self.env.get_curr_cards()), self.action_space, to_char(self.env.get_last_cards()))
s = env.get_state()
s = np.reshape(s, [1, -1])
policy, val = self.sess.run([
self.agents[0].network.valid_policy,
self.agents[0].network.val_pred],
feed_dict={
self.agents[0].network.input: s,
self.agents[0].network.mask: np.reshape(mask, [1, -1])
})
policy = policy[0]
valid_actions = np.take(np.arange(self.a_dim), mask.nonzero())
valid_actions = valid_actions.reshape(-1)
# a = np.random.choice(valid_actions, p=policy)
a = valid_actions[np.argmax(policy)]
# print("taking action: ", self.action_space[a])
return env.step(self.action_space[a])
from network import set_network, create_network
from process_image import img_width
if __name__ == '__main__':
set_network()
main_qn = load_model('./model/model.h5')
main_qn.compile(loss=huber_loss, optimizer=Adam(lr=0.0005))
target_qn = create_network()
memory = Memory(memory_size)
total_step = 0
for episode in range(1, num_episodes + 1):
while True:
state = get_state()
if start_of_episode(state) == 1:
break
step, action, value = 0, 1, 0
do_learn, dead = True, False
state_deque = deque(maxlen=state_deque_size)
target_qn.model.set_weights(main_qn.model.get_weights())
for _ in range(1, max_steps + 1):
step += 1
total_step += 1
# epsilon decay
epsilon = epsilon_stop + (epsilon_start - epsilon_stop) * np.exp(
def rl(q_table):
for episode in range(MAX_EPISODES):
print("episode", episode)
environment = pd.DataFrame(np.zeros((3, 3)))
result = "continue"
# 玩家1操作
state1 = env.get_state(environment)
action1 = choose_action(state1, q_table, environment)
env.get_result(environment, 1, action1)
while True:
# 玩家2操作
state2 = env.get_state(environment)
action2 = choose_action(state2, q_table, environment)
result = env.get_result(environment, 2, action2)
state1_ = env.get_state(environment) # 玩家2完成动作后，就是玩家1的转移后的状态
if result == 'win':
R1, R2 = -1, 1
q_table.loc[state1, action1] += ALPHA * (
R1 + LAMBDA * q_table.iloc[state1_, :].max() -
q_table.loc[state1, action1])
q_table.loc[state2,
action2] += ALPHA * (R2 -
q_table.loc[state2, action2])
break
elif result == 'continue':
R1 = 0
q_table.loc[state1, action1] += ALPHA * (
R1 + LAMBDA * q_table.iloc[state1_, :].max() -
q_table.loc[state1, action1])
else:
R1, R2 = 0.1, 0.1
q_table.loc[state1, action1] += ALPHA * (
R1 + LAMBDA * q_table.iloc[state1_, :].max() -
q_table.loc[state1, action1])
q_table.loc[state2,
action2] += ALPHA * (R2 -
q_table.loc[state2, action2])
break
# 玩家1操作
state1 = env.get_state(environment)
action1 = choose_action(state1, q_table, environment)
result = env.get_result(environment, 1, action1)
state2_ = env.get_state(environment)
if result == 'win':
R1, R2 = 1, -1
q_table.loc[state2, action2] += ALPHA * (
R2 + LAMBDA * q_table.iloc[state2_, :].max() -
q_table.loc[state2, action2])
q_table.loc[state1,
action1] += ALPHA * (R1 -
q_table.loc[state1, action1])
break
elif result == 'continue':
R2 = 0
q_table.loc[state2, action2] += ALPHA * (
R2 + LAMBDA * q_table.iloc[state2_, :].max() -
q_table.loc[state2, action2])
else:
R1, R2 = 0.1, 0.1
q_table.loc[state2, action2] += ALPHA * (
R2 + LAMBDA * q_table.iloc[state2_, :].max() -
q_table.loc[state2, action2])
q_table.loc[state1,
action1] += ALPHA * (R1 -
q_table.loc[state1, action1])
break
return q_table