action = agent.act(state)
img_state, reward, done = game.make_action(action)
if not done:
state_new = img_state
else:
state_new = None
agent.add_transition(state, action, reward, state_new, done)
state = state_new
if learning_step % UPDATE_FREQUENCY == 0:
agent.learn_from_memory()
if learning_step % COPY_FREQUENCY == 0:
updateTarget(targetOps, SESSION)
if done:
print("Epoch %d Train Game %d get %.1f" % (epoch, games_cnt, game.get_total_reward()))
break
if SAVE_MODEL and games_cnt % 10 == 0:
saver.save(SESSION, model_savefile)
print("Saving the network weigths to:", model_savefile)
print("\nTesting...")
test_scores = []
for test_step in range(EPISODES_TO_TEST):
game.reset()
agent.reset_cell_state()
while not game.is_terminared():
state = game.get_state()
action = agent.act(state, train=False)
game.make_action(action)
trainables = tf.trainable_variables()
targetOps = updateTargetGraph(trainables, TAU)
print("Loading model from: ", model_savefile)
saver.restore(SESSION, model_savefile)
##########################################
print("\nTesting...")
test_scores = []
for test_step in range(EPISODES_TO_TEST):
game.reset()
agent.reset_cell_state()
while not game.is_terminared():
state = game.get_state()
action = agent.act(state, train=False)
game.make_action(action)
now_score = game.get_total_reward()
saveScore(now_score)
test_scores.append(now_score)
test_scores = np.array(test_scores)
my_file = open(reward_savefile, 'a') # Name and path of the reward text file
my_file.write("%.1f (±%.1f) min:%.1f max:%.1f\n" %
(test_scores.mean(), test_scores.std(), test_scores.min(),
test_scores.max()))
my_file.close()
img_state, reward, done = game.make_action(action)
if not done:
state_new = img_state
else:
state_new = None
agent.add_transition(state, action, reward, state_new, done)
state = state_new
if learning_step % UPDATE_FREQUENCY == 0:
agent.learn_from_memory()
if learning_step % COPY_FREQUENCY == 0:
updateTarget(targetOps, SESSION)
if done:
print("Epoch %d Train Game %d get %.1f" %
(epoch, games_cnt, game.get_total_reward()))
break
if SAVE_MODEL and games_cnt % 10 == 0:
saver.save(SESSION, model_savefile)
print("Saving the network weigths to:", model_savefile)
print("\nTesting...")
test_scores = []
for test_step in range(EPISODES_TO_TEST):
game.reset()
agent.reset_cell_state()
while not game.is_terminared():
state = game.get_state()
action = agent.act(state, train=False)
game.make_action(action)
s, reward, d = game.make_action(action)
done = game.is_terminared()
if not done:
state_new = preprocess(game.get_state())
else:
state_new = None
agent.add_transition(state, action, reward, state_new, done)
state = state_new
if learning_step % UPDATE_FREQUENCY == 0:
agent.learn_from_memory()
updateTarget(targetOps, SESSION)
if done:
train_scores.append(game.get_total_reward())
train_episodes_finished += 1
game.reset()
agent.reset_cell_state()
state = preprocess(game.get_state())
print("%d training episodes played." % train_episodes_finished)
train_scores = np.array(train_scores)
print(
"Results: mean: %.1f±%.1f," %
(train_scores.mean(), train_scores.std()),
"min: %.1f," % train_scores.min(),
"max: %.1f," % train_scores.max())
print("\nTesting...")