saver.restore(sess, '/home/carl/PycharmProjects/OpenAIUniverseRL/train')
while True:
action_n = [actionSpace[:][action]
for ob in observation_n] # your agent here
observation_n, reward_n, done_n, info = env.step(action_n)
rewardNormalized = np.array([reward_n[0] / 40
]) # to make the max reward closer to 1.0
if observation_n[0]:
pixels_raw = observation_n[0].get("vision")[
84:592, 18:818] # Size 508x800. Will be resized by 0.1
grayscaleImg = ImageProcessing.pre_process_image(
pixels_raw) # Makes a 51x80 'uint8' list
counter += 1
if counter == 1:
motionTracer = ImageProcessing.MotionTracer(
pixels_raw) # create the object
else:
motionTracer.process(pixels_raw) # do stuff to create motion image
if counter > 399: # Ignore the unearned reward while the car still has it's initial momentum at the start
# Add data to the replay memory
# Corresponding Action and Reward gets added one env.step latter than state and q_values
RM.stateArray = np.vstack(
(RM.stateArray, state)) # 'state' comes from previous step
RM.q_valueArray = np.concatenate(
(RM.q_valueArray, Q_values_est),
axis=0) # 'Q_values_est' comes from previous step
RM.actionArray = np.concatenate(
(RM.actionArray, action_as_1D_array), axis=0)
RM.rewardArray = np.concatenate(
(RM.rewardArray, rewardNormalized), axis=0)
# state = Returs 'grayscale' channel [0] and 'grayscale-motion trace' channel [1]
