print('lr:', lr)
print('df:', df)
print('nhl:', nhl)
print('ef:', ef)
# Q-Network, set parameters
QN = DQNAgent(state_size, action_size, discount_factor=df, learning_rate=lr, expl_decay=ef, nhl=nhl, sl_f=1)
batch_size = 16
loss_overall = []
error_avg = []
error = []
states = sim_env.state
for e in range(episodes):
sim_env.reset()
for k in range(training):
states = np.reshape(states, [1, state_size])
action = QN.act(states)
next_state, rewards, overall_err = sim_env.Assign_Cores(action)
next_state = np.reshape(next_state, [1, state_size])
QN.remember(states, action, rewards, next_state)
states = next_state
if len(QN.memory) > batch_size:
QN.replay(batch_size)
loss_overall = np.append(loss_overall, QN.loss_avg/training)
QN.loss_avg = 0
sim_env.reset()
for u in range(testing):
states = np.reshape(states, [1, state_size])
action = QN.act_test(states)
# print('SNR:', sim_env.SNR[-1])
env = wrap_dqn(gym.make('PongNoFrameskip-v4'))
agent = DQNAgent(env=env, num_actions=NUM_ACTIONS, lr=LR, discount=GAMMA)
# Load model
# agent.load_model(weights_file="snaps/model")
# Train agent
agent.train(TRAIN_STEPS, weights_file="snaps/model")
# Evaluate
success = 0
for tr in range(TRIALS):
state = env.reset()
t = 0
acc_r = 0
while True:
env.render()
action = agent.act(state)
state, reward, done, _ = env.step(action)
acc_r += reward
t += 1
if done:
print("Trial {} finished after {} timesteps".format(tr, t))
if acc_r > 0:
success += 1
break
print("Success: %d/%d" % (success, TRIALS))
env.close()
br = tetris.detect_figure(colors)
num += 1
dont_burn_my_cpu.tick(maxfps)
if br == None:
app.gameover = True
if br != None:
if episodes == EPISODES:
app.gameover = True
br = None
continue
action = agent.act(numpy.reshape(tetris.state, [20, 10]), tetris,
dict, br)
state = tetris.generate_state_based_on_action_and_figure(
dict, action, br)
reward1 = app.score
for i in range(action[2]):
app.rotate_stone()
if action[1] == 0:
app.move(-1 * action[0])
elif action[1] == 1:
app.move(1 * action[0])
app.insta_drop()