game = GAME()
game.set_timer(100)
network = Network(n_output, lr_rate)
agent = Agent(network, gamma, batch_size, n_action, input_dims, mem_size)
scores = []
avg_rewards = []
score_max = 0
n_total = 0
for n in range(n_games):
if not n == 0:
scores.append(score)
avg_rewards.append(total_reward / n_moves)
agent.learn(n_total, n_total + n_moves)
score = 0
game.score = 0
total_reward = 0
n_moves = 0
n_total = n_moves+n_total
done = False
observation = game.get_states()
while not done:
action = agent.choose_action(observation)
if n < -1:
game.draw()
action = int(msvcrt.getch().decode("utf-8"))
observation_, reward, done, score = game.spin_once(action, n_moves)
agent.store_data(observation, action, reward, observation_, done)
score_record = []
score_record_step = []
count_record = []
count_record_step = []
time_record = []
time_record_step = []
for i in range(800):
done = False
score = 0
obs = env.get_init_state()
# 没分配完
while not done:
act = MECSnet.choose_action(obs)
new_state, reward, done = env.step(act)
MECSnet.remember(obs, act, reward, new_state, int(done))
MECSnet.learn()
score += reward
obs = new_state
# print('reward is： {}'.format(reward))
# 本轮的reward追加到list中
score_record.append(score)
# print('episode ', i, 'score %.2f' % score, '100 game average %.2f' % np.mean(score[-100:]))
print('episode ', i, 'score %.2f' % score, " wrong: ", env.count_wrong)
count_record.append(1 - env.count_wrong / num_task)
time_record.append(env.time)
if i % 25 == 0:
# UEnet.save_models()
MECSnet.save_models()
score_record_step.append(np.mean(score_record))
count_record_step.append(np.mean(count_record))