def main():
experiment = 'model-builder-v0' #specify environments here
env = gym.make(experiment)
#steps= env.spec.timestep_limit #steps per episode
steps = 20
assert isinstance(env.observation_space,
Box), "observation space must be continuous"
assert isinstance(env.action_space, Box), "action space must be continuous"
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm)
exploration_noise = OUNoise(env.action_space.shape[0])
counter = 0
reward_per_episode = 0
total_reward = 0
num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
print("Number of States:", num_states)
print("Number of Actions:", num_actions)
print("Number of Steps per episode:", steps)
#saving reward:
reward_st = np.array([0])
for i in range(episodes):
print("==== Starting episode no:", i, "====", "\n")
observation = env.reset()
reward_per_episode = 0
for t in range(steps):
#rendering environmet (optional)
env.render()
x = observation
action = agent.evaluate_actor(np.reshape(x, [1, 300, 300, 2]))
noise = exploration_noise.noise()
action = action[
0] + noise #Select action according to current policy and exploration noise
print("Action at step", t, " :", action, "\n")
observation, reward, done, info = env.step(action)
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x, observation, action, reward, done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode += reward
counter += 1
#check if episode ends:
if (done or (t == steps - 1)):
print('EPISODE: ', i, ' Steps: ', t, ' Total Reward: ',
reward_per_episode)
print("Printing reward to file")
exploration_noise.reset(
) #reinitializing random noise for action exploration
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, newline="\n")
print('\n\n')
break
total_reward += reward_per_episode
print("Average reward per episode {}".format(total_reward / episodes))
def main():
experiment= 'InvertedPendulum-v1' #specify environments here
env= gym.make(experiment)
steps= env.spec.timestep_limit #steps per episode
assert isinstance(env.observation_space, Box), "observation space must be continuous"
assert isinstance(env.action_space, Box), "action space must be continuous"
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm)
exploration_noise = OUNoise(env.action_space.shape[0])
counter=0
reward_per_episode = 0
total_reward=0
num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
print "Number of States:", num_states
print "Number of Actions:", num_actions
print "Number of Steps per episode:", steps
#saving reward:
reward_st = np.array([0])
for i in xrange(episodes):
print "==== Starting episode no:",i,"====","\n"
observation = env.reset()
reward_per_episode = 0
for t in xrange(steps):
#rendering environmet (optional)
env.render()
x = observation
action = agent.evaluate_actor(np.reshape(x,[1,num_states]))
noise = exploration_noise.noise()
action = action[0] + noise #Select action according to current policy and exploration noise
print "Action at step", t ," :",action,"\n"
observation,reward,done,info=env.step(action)
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x,observation,action,reward,done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode+=reward
counter+=1
#check if episode ends:
if (done or (t == steps-1)):
print 'EPISODE: ',i,' Steps: ',t,' Total Reward: ',reward_per_episode
print "Printing reward to file"
exploration_noise.reset() #reinitializing random noise for action exploration
reward_st = np.append(reward_st,reward_per_episode)
np.savetxt('episode_reward.txt',reward_st, newline="\n")
print '\n\n'
break
total_reward+=reward_per_episode
print "Average reward per episode {}".format(total_reward / episodes)
def main():
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm, CA_OBS_SPACE, CA_ACTION_SPACE,
CA_ACTION_BOUND)
exploration_noise = OUNoise(CA_ACTION_SPACE)
counter = 0
reward_per_episode = 0
total_reward = 0
num_states = CA_OBS_SPACE
num_actions = CA_ACTION_SPACE
print "Number of States:", num_states
print "Number of Actions:", num_actions
print "Number of Steps per episode:", steps
#saving reward:
reward_st = np.array([0])
for i in xrange(episodes):
print "==== Starting episode no:", i, "====", "\n"
# observation = env.reset()
observation = ca_reset()
reward_per_episode = 0
for t in xrange(steps):
#rendering environmet (optional)
# env.render()
x = observation
action = agent.evaluate_actor(np.reshape(x, [1, num_states]))
noise = exploration_noise.noise()
action = action[
0] + noise #Select action according to current policy and exploration noise
print "Action at step", t, " :", action, "\n"
# observation,reward,done,info=env.step(action)
observation, reward, done, info = ca_step(action)
print x, observation, action, reward, done
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x, observation, action, reward, done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode += reward
counter += 1
#check if episode ends:
if (done or (t == steps - 1)):
print 'EPISODE: ', i, ' Steps: ', t, ' Total Reward: ', reward_per_episode
print "Printing reward to file"
exploration_noise.reset(
) #reinitializing random noise for action exploration
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, newline="\n")
print '\n\n'
break
total_reward += reward_per_episode
print "Average reward per episode {}".format(total_reward / episodes)
def main():
'''main function'''
# Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(ACTORNET_PRE_TRAINED, STATENET_PRE_TRAINED)
exploration_noise = OUNoise(ACTION_DIM)
# saving reward:
reward_st = np.array([0])
img_server = ImgServer()
img_server.wait_for_connect()
observe_t_img = img_server.receive_img()
observe_t_data = convert_img2data(observe_t_img)
actor_t = agent.evaluate_actor(observe_t_data)
noise = exploration_noise.noise()
actor_t = actor_t[0] + noise
img_server.send_actor_cmd(actor_t)
observe_t_1_img = observe_t_img
actor_t_1 = actor_t
img_server.close_connect()
index = 1
while True:
observe_t_img = observe_t_1_img
actor_t = actor_t_1
img_server.wait_for_connect()
observe_t_1_img = img_server.receive_img()
observe_t_1_data = convert_img2data(observe_t_1_img)
actor_t_1 = agent.evaluate_actor(observe_t_1_data)
noise = exploration_noise.noise()
actor_t_1 = actor_t_1[0] + noise
cost = compute_cost(observe_t_img)
agent.add_experience(observe_t_img, observe_t_1_img, actor_t, cost, index)
if index > 32:
agent.train()
img_server.send_actor_cmd(actor_t_1)
img_server.close_connect()
index = index + 1
def main():
env = Env(19997)
steps= 10000
num_states = 59
num_actions = 3
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm)
exploration_noise = OUNoise(num_actions)
counter=0
reward_per_episode = 0
total_reward=0
reward_st = np.array([0])
agent.actor_net.load_actor(os.getcwd() + '/weights/actor/model.ckpt')
agent.critic_net.load_critic(os.getcwd() + '/weights/critic/model.ckpt')
for i in range(episodes):
# print "==== Starting episode no:",i,"====","\n"
observation = env.reset()
done =False
reward_per_episode = 0
for t in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x,[1,num_states]))
noise = exploration_noise.noise()
action = action[0] + noise #Select action according to current policy and exploration noise
for i in range(num_actions):
if action[i] > 1.0:
action[i] = 1.0
if action[i] < -1.0:
action[i] = -1.0
observation,reward,done = env.step(action)
print("reward:", reward, "\n")
agent.add_experience(x,observation,action,reward,done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode+=reward
counter+=1
#check if episode ends:
if (done or (t == steps-1)):
print('Episode',i,'Steps: ',t,'Episode Reward:',reward_per_episode)
exploration_noise.reset()
reward_st = np.append(reward_st,reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, newline="\n")
agent.actor_net.save_actor(os.getcwd() + '/weights/actor/model.ckpt')
agent.critic_net.save_critic(os.getcwd() + '/weights/critic/model.ckpt')
break
total_reward+=reward_per_episode
def main():
env= Env(19997)
steps = 300
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm)
exploration_noise = OUNoise(2)
counter = 0
reward_per_episode = 0.
num_states = 32*16
num_actions = 2
#saving reward:
reward_st = np.array([0])
for i in range(episodes):
print ("==== Starting episode no:",str(i),"====","\n")
observation = env.reset()
reward_per_episode = 0
for t in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x,[1,num_states]))
noise = exploration_noise.noise()
action = action[0] + noise
observation,reward,done=env.step(action,t)
agent.add_experience(x,observation,action,reward,done)
if counter > 64:
agent.train()
reward_per_episode+=reward
counter+=1
#check if episode ends:
if (done):
print ('EPISODE: ',str(i),' Steps: ',str(t),' Total Reward: ',str(reward_per_episode))
exploration_noise.reset() #reinitializing random noise for action exploration
reward_st = np.append(reward_st,reward_per_episode)
np.savetxt('episode_reward.txt',reward_st, newline="\n")
agent.actor_net.save_actor('/home/lee/Projects/Tracking/RL/weights/actor/model.ckpt')
agent.critic_net.save_critic('/home/lee/Projects/Tracking/RL/weights/critic/model.ckpt')
print ('\n\n')
break
def main():
env = Env(20000)
steps = 50
agent = DDPG(env, is_batch_norm)
counter = 0
exploration_noise = OUNoise(2)
reward_per_episode = 0
num_states = 96 * 4 + 4
num_actions = 2
reward_st = np.array([0])
agent.actor_net.load_actor(
'/home/myounghoe/ddpgtf/norepeat_target_2action_scale2/weights/actor/model.ckpt'
)
agent.critic_net.load_critic(
'/home/myounghoe/ddpgtf/norepeat_target_2action_scale2/weights/critic/model.ckpt'
)
for i in range(episodes):
print("==== Starting episode no:", str(i), "====", "\n")
observation = env.reset()
reward_per_episode = 0
for t in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x, [1, num_states]))
noise = exploration_noise.noise()
action = action[0] # + noise
action = np.array([-1.0, 0.0])
observation, reward, done = env.step(action, t)
reward_per_episode += reward
counter += 1
#check if episode ends:
if (done):
print('EPISODE: ', str(i), ' Steps: ', str(t),
' Total Reward: ', str(reward_per_episode))
# print "Printing reward to file"
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('test_reward.txt', reward_st, newline="\n")
print('\n\n')
break
def main():
with tf.Graph().as_default():
agent = DDPG(number_of_states, number_of_actions)
reward_per_time_step = 0
RM = ReplayMemory(100000)
for e in range(episodes):
print('Begin Episode number', e)
# For Loop του αλγορίθμου για ένα επισόδειο
for t in range(steps):
if t == 0:
current_moisture = 0.01
else:
current_moisture = RM.replay_memory[-1][0][0] # y(t)
current_state = np.array([current_moisture, y_set]) # s
current_state_true = current_state.reshape(1, 2)
action = agent.evaluate_actor(current_state_true)[0][
0] # Δίνει το action , α(t)
print(action)
T = np.linspace(t, t + 1)
next_moisture = output(sys, T, action,
current_moisture) # y(t+1)
next_state = np.array([next_moisture, y_set]) # s'
current_reward = reward(agent.model_train()[-1]) # r
# print("this thing =", agent.model_train()[-1])
reward_per_time_step += current_reward # Συνολικό reward
# print(reward_per_time_step)
RM.add_experience(current_state, next_state, action,
current_reward)
agent.model_train()
def main():
env = Env(20000)
steps= 10000
num_states = 59
num_actions = 3
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm)
counter=0
reward_per_episode = 0
total_reward=0
reward_st = np.array([0])
agent.actor_net.load_actor(os.getcwd() + '/weights/actor/model.ckpt')
agent.critic_net.load_critic(os.getcwd() + '/weights/critic/model.ckpt')
for i in range(episodes):
# print "==== Starting episode no:",i,"====","\n"
observation = env.reset()
done =False
reward_per_episode = 0
for t in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x,[1,num_states]))
action = action[0]
observation,reward,done = env.step(action)
reward_per_episode+=reward
counter+=1
#check if episode ends:
if (done or (t == steps-1)):
print('Episode',i,'Steps: ',t,'Episode Reward:',reward_per_episode)
reward_st = np.append(reward_st,reward_per_episode)
# np.savetxt('episode_reward.txt', reward_st, newline="\n")
break
total_reward+=reward_per_episode
def main():
sess = tf.Session()
setting.load_data(setting.currency, train_test_data.file_list,
train_test_data.test_file)
agent = DDPG(sess, CURRENCY, CHART, TIMELINE, LENGTH)
counter = 0
reward_for_episode = 0
total_reward = 0
epsilon = 1.0 # parameter defining ratio between random action and DQN decision
time_step = 0 # frame number
# saving reward
reward_st = np.array([0])
saver = tf.train.Saver(tf.global_variables())
ckpt = tf.train.get_checkpoint_state('./trade_model')
if ckpt and tf.train.checkpoint_exists(ckpt.model_checkpoint_path):
print('model has been loaded successfully!')
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print('start new progress.')
sess.run(tf.global_variables_initializer())
for idx in range(MAX_EPISODE):
terminal = False
print('Starting episode no: %d' % idx)
state = setting.reset()
reward_for_episode = 0
step_on_episode = 0
while not terminal:
present_state = state
if np.random.rand() < epsilon:
selected_currency = np.random.choice(CURRENCY)
ratio = 2 * (np.random.rand() - 0.5)
action = setting.action_value(CURRENCY, selected_currency,
ratio)
else:
action = agent.evaluate_actor(present_state)
if idx > OBSERVE:
epsilon -= 1 / 50000
state, reward, terminal, _ = setting.step(action)
# add s_t, s_(t+1), action, reward to experience memory
agent.add_experience(present_state, state, action, reward,
terminal)
# train critic and actor network
if time_step > 2000 and time_step % TRAIN_INTERVAL == 0:
agent.train()
reward_for_episode += reward
time_step += 1
step_on_episode += 1
# check if episode ends
print('at %s, EPISODE: %d, Steps: %d, Reward: %d' %
(str(datetime.datetime.now()), idx, step_on_episode,
reward_for_episode))
reward_st = np.append(reward_st, reward_for_episode)
if idx % 500 == 0 and idx != 0:
saver.save(sess,
'trade_model/actor_critic_network.ckpt',
global_step=time_step)
total_reward += reward_for_episode
print('Average reward per episode: {}'.format(total_reward / MAX_EPISODE))
def main():
'''
In this file, we first load the system state parameter from the .mat files, then for each
each slot, we observe the state parameter and make the action. Then, we save this state-actor
record into the memory for the latter train. Finally, the system convert into te next ecopids.
'''
#load the state parameter form .mat file
task_size = sio.loadmat('./data/data')['input_data_size'] #load the task size
CPU_density = sio.loadmat('./data/data')['input_CPU_density'] #load the required CPU cycles of each task bit
task_delay_re = sio.loadmat('./data/data')['input_task_delay_re'] #load the maximum toleration delay of each task
task_gain = sio.loadmat('./data/data')['input_task_gain'] #load the gain of each task
user_power = sio.loadmat('./data/data')['input_user_power'] #load the transmit power of each user
user_chan_gain = sio.loadmat('./data/data')['input_user_chan_gain'] #load the wireless channel gain of each user
bs_capacity = sio.loadmat('./data/data')['input_bs_capacity'] #load the computing capacity of each base station
#set the number of users in these base station
bs_1_user_num = 10
bs_2_user_num = 20
bs_3_user_num = 10
#set the wireless channel nosie, channel bandiwidth, transmission rate of wired connection,
chan_noise = 10**(-8)
chan_band = 10**6
wired_rate = 10
#set the length of time slot
slot_len = 10000
#Set the record number in the replay buffer, the total reward, the reward record of the whole time slots
counter = 0
total_reward = 0
reward_st = np.array([0])
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
num_states, num_actions = len(task_size[:,1]) * 7, len(task_size[:,1])
agent = DDPG(num_states, num_actions, is_batch_norm)
#set the explore nosie to guarantee the algrithm's optimal performance
exploration_noise = OUNoise(1)
#travel each slot, and make the action decision
for i in range(slot_len):
print ("==== Starting episode no:",i,"====","\n")
current_state = np.hstack((task_size[:,i], CPU_density[:,i], task_delay_re[:,i], task_gain[:,i],\
user_power[:,0], user_chan_gain[:,i],bs_capacity[:,i])) #obtain the current system state
current_state = np.reshape(current_state, [1, -1])
actor_input = current_state #set the input of actor network
actor_output = agent.evaluate_actor(actor_input) #predict the action in this slot
noise = exploration_noise.noise() #obtain the noise added in the action
action = actor_output[0] + noise #Select action according to current policy and exploration noise
# print ("Action at slot", i ," :",action,"\n")
reward = 1#fuction(action,current_state) #obtain the reward in this slot
next_state = np.hstack((task_size[:,i+1], CPU_density[:,i+1], task_delay_re[:,i+1], task_gain[:,i+1], user_power[:,0],\
user_chan_gain[:,i+1], bs_capacity[:,i+1])) #obtain the system state in the next slot
next_state = np.reshape(next_state, [1, -1])
agent.add_experience(current_state, next_state, action, reward) #add s_t,s_t+1,action,reward to experience memory
#train critic and actor network
if counter > 64:
agent.train()
counter+=1
# print ('EPISODE: ',i,'Reward: ',reward)
reward_st = np.append(reward_st,reward)
np.savetxt('episode_reward.txt',reward_st, newline="\n")
total_reward+=reward
print ("Average reward per episode {}".format(total_reward / slot_len))
def s2l():
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
num_states = feature_size #num_states = env.observation_space.shape[0]
num_actions = num_controls
print ("Number of States:", num_states)
print ("Number of Actions:", num_actions)
action_space_high=[1.5] #[0.0,0.0,0.0]
action_space_low=[0.03] #[0.5,0.5,0.5]
print ("Action space highest values", action_space_high)
print ("Action space lowest values:", action_space_low)
robot=RoboControl()
#while True:
# #robot.check()
# robot.publish_control([1])
# robot.reset()
agent = DDPG(is_batch_norm,num_states,num_actions,action_space_high,action_space_low)
exploration_noise = OUNoise(num_actions)
counter=0
total_reward=0
print ("Number of Rollouts per episode:", num_rollouts)
print ("Number of Steps per roll out:", steps)
reward_st = np.array([0]) #saving reward
eval_metric_st= np.array([0])
reward_st_all = np.array([0]) #saving reward after every step
activity_obj=Vid_Feature()
demo_vid_array=demo_array_extractor(demo_folder)
demo_features=activity_obj.feature_extractor(demo_vid_array)
frame_obj=Frame_Feature()
#camera_obj= Camera()
camera_obj= CameraSub()
for episode in range(num_episodes):
print ("==== Starting episode no:",episode,"====","\n")
robot.reset() # Reset env in the begining of each episode
obs_img=camera_obj.camera_subscribe() # Get the observation
#obs_img=np.array(misc.imresize(obs_img,[112,112,3]))
observation =np.array(frame_obj.frame_feature_extractor(obs_img))
observation=observation.reshape(-1)
reward_per_episode = 0
for t in range(num_rollouts):
reward_per_rollout=0
vid_robo_=[]
for i in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x,[1,num_states]))
noise = exploration_noise.noise()
action = action[0] + noise #Select action according to current policy and exploration noise
print ('Action at episode-',episode,'rollout-',t, 'step-', i ," :",action)
robot.publish_control(action)
obs_robo=camera_obj.camera_subscribe() # Get the observation
#obs_robo=misc.imresize(obs_robo,[112,112,3])
vid_robo_.append(obs_robo)
observation=np.array(frame_obj.frame_feature_extractor(np.array(obs_robo)))
observation=observation.reshape(-1)
#pasue()
if(i==15):
vid_robo=np.array(vid_robo_)
robo_features=activity_obj.feature_extractor(vid_robo)
reward=-(distance(demo_features,robo_features))
reward=np.array(reward)
print('reward: ',reward)
else:
reward=0
reward=np.array(reward)
print('reward: ',reward)
# Storing reward after every rollout
reward_st_all = np.append(reward_st_all,reward)
np.savetxt('reward_all.txt',reward_st_all, newline="\n")
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x,observation,action,reward,False)
reward_per_rollout+=reward
counter+=1
#train critic and actor network
if counter > start_training:
agent.train()
print ('\n\n')
#Saving policy
if ((episode%100)==0 and t==num_rollouts-1):
print('saving policy...........................!')
agent.save_actor(episode)
reward_per_episode+=reward_per_rollout
#check if episode ends:
print ('EPISODE: ',episode,' Total Reward: ',reward_per_episode)
print ("Printing reward to file")
exploration_noise.reset() #reinitializing random noise for action exploration
reward_st = np.append(reward_st,reward_per_episode)
np.savetxt('episode_reward.txt',reward_st, fmt='%f', newline="\n")
print ('\n\n')
total_reward+=reward_per_episode
print ("Average reward per episode {}".format(total_reward / num_episodes))
def main():
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG()
exploration_noise = OUNoise(env.action_space.shape[0])
counter=0
reward_per_episode = 0
total_reward=0
#saving reward:
reward_st = np.array([0])
# network setup
s = socket.socket() # Create a socket object
#host = socket.gethostname() # Get local machine name
host = '' # Get local machine name
port = 21567 # Reserve a port for your service.
s.bind((host, port))
s.listen(5)
imgorigin_t = np.zeros(300,400)
imgorigin_t_1 = np.zeros(300,400)
actor_t = np.zeros(6)
actor_t_1 =np.zeros(6)
index = 0
#the first time
c, addr = s.accept() # Establish connection with client.
print ('Got connection from'), addr
print ("Receiving...")
l = c.recv(1024)
f = open('temp.tif','wb')
while (l):
f.write(l)
l = c.recv(1024)
f.close()
print ("Done Receiving")
imgorigin_t = np.array(Image.open('temp.tif'))
tempimg = imgorigin_t[np.newaxis,:,:,np.newaxis]
tempimg = tempimg.transpose([0,2,1,3])
test_pred = agent.evaluate_actor(tempimg)
action_t = test_pred[0]
print action_t
str_buf = ''
str_buf = str_buf+str(action_t[0,0])+" "
str_buf = str_buf+str(action_t[0,1])+" "
str_buf = str_buf+str(action_t[0,2])+" "
str_buf = str_buf+str(action_t[0,3])+" "
str_buf = str_buf+str(action_t[0,4])+" "
str_buf = str_buf+str(action_t[0,5])+" "
imgorigin_t_1 = imgorigin_t
actor_t_1 = actor_t
c.send(str_buf)
c.close()
index =1
while True:
#update imgorigin_t and actor_t
imgorigin_t = img_origin_t_1
actor_t = actor_t_1
c, addr = s.accept() # Establish connection with client.
print ('Got connection from'), addr
print ("Receiving...")
l = c.recv(1024)
f = open('temp.tif','wb')
while (l):
f.write(l)
l = c.recv(1024)
f.close()
print ("Done Receiving")
imgorigin_t_1 = np.array(Image.open('temp.tif'))
tempimg = imgorigin_t_1[np.newaxis,:,:,np.newaxis]
tempimg = tempimg.transpose([0,2,1,3])
test_pred = agent.evaluate_actor(tempimg)
action_t_1 = test_pred[0]
print action_t_1
reward = compute_reward(imgorigin_t_1)
agent.add_experience(imgorigin_t,imgorigin_t_1,action_t,reward,index)
if index > 32:
agent.train()
str_buf = ''
str_buf = str_buf+str(action_t_1[0,0])+" "
str_buf = str_buf+str(action_t_1[0,1])+" "
str_buf = str_buf+str(action_t_1[0,2])+" "
str_buf = str_buf+str(action_t_1[0,3])+" "
str_buf = str_buf+str(action_t_1[0,4])+" "
str_buf = str_buf+str(action_t_1[0,5])+" "
c.send(str_buf)
print("send action finished!")
c.close()
index = index+1
def main():
experiment = 'quadruped-robot-v0' #specify environments here
backupNameFile = "quadruped_robot_0"
backupPathFile = "storage/" + backupNameFile
bFullPath = os.path.join(
os.path.split(os.path.abspath(__file__))[0], backupPathFile)
env = gym.make(experiment)
steps = env.spec.timestep_limit #steps per episode
assert isinstance(env.observation_space,
Box), "observation space must be continuous"
assert isinstance(env.action_space, Box), "action space must be continuous"
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
global agent
agent = DDPG(env, is_batch_norm)
exploration_noise = OUNoise(env.action_space.shape[0])
counter = 0
reward_per_episode = 0
total_reward = 0
num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
print("Number of States:", num_states)
print("Number of Actions:", num_actions)
print("Number of Steps per episode:", steps)
#saving reward:
reward_st = np.array([0])
for i in range(episodes):
print("==== Starting episode no:", i, "====", "\n")
observation = env.reset()
reward_per_episode = 0
for t in range(steps):
#rendering environmet (optional)
env.render()
x = observation
action = agent.evaluate_actor(np.reshape(x, [1, num_states]))
noise = exploration_noise.noise()
action = action[
0] + noise #Select action according to current policy and exploration noise
# print ("Action at step", t ," :",action,"\n")
observation, reward, done, info = env.step(action)
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x, observation, action, reward, done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode += reward
counter += 1
#check if episode ends:
if (done or (t == steps - 1)):
# print ('EPISODE: ',i,' Steps: ',t,' Total Reward: ',reward_per_episode)
# print ("Printing reward to file")
exploration_noise.reset(
) #reinitializing random noise for action exploration
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, newline="\n")
print('\n\n')
break
# Save some episodes
# print(episodes)
# if (episodes == 10):
# with open(bFullPath+"_EP_"+episodes+".pkl", 'wb') as file:
# pickle.dump(agent, file)
# pickle.dump_session(bFullPath+"_EP_"+episodes+".pkl")
# print ('SAVE EPISODE ',episodes)
# break;
total_reward += reward_per_episode
print("Average reward per episode {}".format(total_reward / episodes))
def main():
experiment= 'InvertedPendulum-v1'
env= gym.make(experiment)
assert isinstance(env.observation_space, Box), "observation space must be continuous"
assert isinstance(env.action_space, Box), "action space must be continuous"
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env)
exploration_noise = OUNoise(env.action_space.shape[0])
counter=0
total_reward=0
num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
#saving reward:
reward_st = np.array([0])
for i in xrange(episodes):
observation = env.reset()
reward_per_episode = 0
for t in xrange(steps):
#rendering environmet (optional)
#env.render()
x = observation
#select action using actor network model
action = agent.evaluate_actor(np.reshape(x,[num_actions,num_states]))
noise = exploration_noise.noise()
action = action[0] + noise
print 'Agent.Action :',action
print '\n'
print '\n'
observation,reward,done,[]=env.step(action)
#add s_t,s_t+1,action,reward to experience memeroy
agent.add_experience(x,observation,action,reward,done)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode+=reward
counter+=1
#check if episode ends:
if done:
print 'EPISODE: ',i,' Steps: ',t,' Total Reward: ',reward_per_episode
exploration_noise.reset()
reward_st = np.append(reward_st,reward_per_episode)
np.savetxt('episode_reward.txt',reward_st, newline="\n")
print '\n'
print '\n'
break
total_reward+=reward_per_episode
print "Average reward per episode {}".format(total_reward / episodes)
class PlayAgent:
def __init__(self, ip="127.0.0.1", id=28888): # ip 는 서버 주소.id 는 agent 식별자
self.ar = ClientActionRobotJava(ip)
self.se = GameStateExtractor() # ㅎ믇
self.tp = TrajectoryPlanner() # 궤적 계산 모듈
self.firstShot = True
self.solved = [] # clear 한 레벨은 1, 아닌 레벨은 0의 값을 가진다
self.currentLevel = -1
self.failedCounter = 0
self.id = id
self.width = 840 # 게임 화면 너비
self.height = 480 # 게임 화면 높이
# 게임 화면 스크린샷(=state) 의 크기. 즉. state의 개수는
# 화면의 픽셀 수와 같다.
# [height, width, RGB 3 channels]
self.num_states = [self.height, self.width, 3]
# Action space 정의
# [거리(0~90px), 각도(0~90degree), tapTime(0~5000ms)]
self.num_actions = 3
self.action_space_high = [90, 75, 50]
self.action_space_low = [0, 0, 0]
self.noise_mean = [20, -20, 0]
self.noise_sigma = [10, 30, 20]
self.ddpg = DDPG(self.num_states, self.num_actions,
self.action_space_high, self.action_space_low, is_batch_norm)
def getNextLevel(self): # 다음 레벨을 얻어온다
level = 0
unsolved = False
for i in range(len(self.solved)):
if self.solved[i] == 0:
unsolved = True
level = i + 1
if level <= self.currentLevel and self.currentLevel < len(self.solved):
continue
else:
return level
if unsolved:
return level
level = (self.currentLevel + 1) % len(self.solved)
if level == 0:
level = len(self.solved)
return level
def checkMyScore(self):
scores = self.ar.checkMyScore() # 현재 점수 확인
level = 1
for s in scores: # 각 level 별 점수 확인
print "||\tlevel %d score is : %d\t||" % (level, s)
if s > 0:
self.solved[level - 1] = 1
level += 1
def getScreenBuffer(self, buffer, width=840, height=480):
"""
현재 게임플레이 스크린샷을 받아온다.
RGB 별로 따로 저장한다.
"""
print "## Get ScreenBuffer"
# returnBuffer's size = (480, 840, 3)
returnBuffer = np.zeros((height, width, 3))
for i in range(height):
for j in range(width):
RGB = buffer.getRGB(j, i)
returnBuffer[i, j, 0] = RGB & 0x0000ff
returnBuffer[i, j, 1] = RGB & 0x00ff00
returnBuffer[i, j, 2] = RGB & 0xff0000
print "## Return ScreenBuffer"
return returnBuffer
def shoot(self, action):
"""
새를 쏘고,
쏜 후의 상태를 반환한다.
"""
# 새총 detection
screenshot = self.ar.doScreenShot()
vision = Vision(screenshot)
sling = vision.findSlingshotMBR()
# 현재 게임 state
pigs = vision.findPigsMBR()
state = self.ar.checkState()
# 새총이 감지되면 플레이하고, 아니라면 스킵
if sling != None:
# 맵에 돼지가 존재하면 임의로 한 마리를 타겟으로 잡고 쏜다.
if len(pigs) != 0:
refPoint = self.tp.getReferencePoint(sling)
print "## Ref Sling Point : ", refPoint
# DDPG 로부터 취할 action을 받아온다
releaseDistance = action[0]
releaseAngle = action[1]
tapTime = action[2]
print "## Release Distance : ", releaseDistance
print "## Release Angle : ", releaseAngle
self.ar.fullyZoomOut()
screenshot = self.ar.doScreenShot()
vision = Vision(screenshot)
_sling = vision.findSlingshotMBR() # zoom out 했을 때 감지된 새총.
if _sling != None:
# zoom out 하지 않았을 때의 새총 위치와 zoom out 한 후의 새총 위치의 차이를 구하여
# 너무 차이가 난다면, 쏘지 않고 다시 screenshot 을 찍어 분석하도록 함
scale_diff = (sling.width - _sling.width) ** 2 + \
(sling.height - _sling.height) ** 2
if scale_diff < 25:
self.ar.shoot(int(refPoint.x), int(refPoint.y), int(
releaseDistance), int(releaseAngle), 0, int(tapTime), True)
print "## Shooting is Done"
state = self.ar.checkState()
if state == state.PLAYING:
self.firstShot = False
else:
print "## Scale is changed. So sling can not execute the shot and will re-segment the image"
else:
print "## No sling was detected. So agent can not execute the shot and will re-segment the image"
return state
def ddpg_run(self):
"""
DDPG algorithm 을 raw pixel data(screenshot)에 대해서 돌린다
"""
info = self.ar.configure(ClientActionRobot.intToByteArray(self.id))
self.solved = np.zeros(info[2])
self.checkMyScore()
print "## current level : %d" % self.currentLevel
# DDPG
# random 하게 critic, actor, target critic net, target actor net 을 초기화하고
# experience memory 도 deque 로 초기화 한다
exploration_noise = OUNoise(
self.num_actions, self.noise_mean, self.noise_sigma)
counter = 1
reward_per_episode = 0 # episode는 한 판을 의미.
total_reward = 0
print "# of States : ", self.num_states
print "# of Actions : ", self.num_actions
# reward 저장
reward_st = np.array([0])
# parameter 로 정한 episode 수 만큼 training 학습 진행
for i in xrange(episodes):
# 다음 레벨 받아오기
self.currentLevel = self.getNextLevel()
# 받아온 레벨이 1~3 이면 해당 레벨 로드, 아니면 1로 초기화 후 로드
if self.currentLevel < 4:
self.ar.loadLevel(self.currentLevel)
else:
self.currentLevel = 1
self.ar.loadLevel(self.currentLevel)
prevscore = 0
reward_per_episode = 0
steps = 0
print "======== Starting Episode No : ", (i + 1), "========", "\n"
# 하나의 episode 에 대한 루프
while True:
# 게임 플레이 screenshot 가져오기
screenshot = self.ar.doScreenShot()
x = self.getScreenBuffer(screenshot, self.width, self.height)
# actor evaluation 을 통해서 다음에 취할 action 을 얻는다
action = self.ddpg.evaluate_actor(np.reshape(
x, [1, self.num_states[0], self.num_states[1], self.num_states[2]]))
print "## Get Action from network!! : ", action
action = action[0]
noise = exploration_noise.noise()
# action 을 현재의 policy 에 따라 정하되,
# epsilon(noise) 수치 정도에 따라 실험적인 action을
# stochastic 하게 취하도록 한다.
action = action + noise
print action
# distance 가 음수이면 양수로 뒤집어준다.
action[0] = action[0] if action[0] > self.action_space_low[0] else -action[0]
# distance 가 최대 범위를 넘어서면 최대 범위로 설정한다.
action[0] = action[0] if action[0] < self.action_space_high[0] else self.action_space_high[0]
# 각도의 경우에도 마찬가지 처리를 해준다
action[1] = action[1] if action[1] > self.action_space_low[1] else -action[1]
action[1] = action[1] if action[1] < self.action_space_high[1] else self.action_space_high[1]
# tap time 도 마찬가지
action[2] = action[2] if action[2] > self.action_space_low[2] else -action[2]
action[2] = action[2] if action[2] < self.action_space_low[2] else self.action_space_high[2]
print "## Action at step ", steps, " :", action, "\n"
# 쏘고나서 점수가 안정화 될 때까지 조금 기다리는 로직이 들어있다
state = self.shoot(action)
if state == state.WON or state == state.LOST:
# episode 가 끝나면( 한 레벨이 끝나면 )
print "## Episode End"
screenshot = self.ar.doScreenShot()
observation = self.getScreenBuffer(
screenshot, self.width, self.height)
# 이기면 reward를 받고 지면 받지 않는다.
if state == state.WON:
score = self.se.getScoreEndGame(screenshot)
# 현재 episode 에서 얻은 점수를 1000으로 나눈 값을 reward 로 사용
reward = (score - prevscore) / 1000.0
else:
reward = 0.00
self.currentLevel = self.currentLevel
self.firstShot = True # episode 가 끝나면 first shot 초기화
done = True # episode done 처리
# experience memory 에
# s(t), s(t + 1), action, reward 를 저장한다
print "######## SCORE : ", score
print "######## REWARD : ", reward
# x = state(screenBuffer) at t
# obervation = state(screenBuffer) at (t + 1)
self.ddpg.add_experience(
x, observation, action, reward, done)
# critic network 와 actor network 학습
# 정해둔 step 이상 진행됐을 경우부터 학습을 시작하도록 한다.
# experience 를 충분히 경험해야 하기 때문.
if counter > TRAIN_STEP:
self.ddpg.train()
counter += 1
steps += 1
print "==== EPISODE: ", i, ' Steps: ', steps, ' Total Reward: ', reward_per_episode
print "Writing reward info into file..."
exploration_noise.reset()
# reward_st 는 배열이다.
# 마지막 원소에 해당 판에서 얻은 총 점수를 기록하고
# 파일로 내보낸다
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt("episodes_reward.txt", reward_st, newline="\n")
print "\n\n"
break
elif state == state.PLAYING: # PLAYING 상태일 때
screenshot = self.ar.doScreenShot()
vision = Vision(screenshot)
sling = vision.findSlingshotMBR()
while sling == None and self.ar.checkState() == state.PLAYING:
print "## No slingshot was detected. Please remove pop up or zoom out"
self.ar.fullyZoomOut()
screenshot = self.ar.doScreenShot()
# S(t + 1) 을 얻는다
observation = self.getScreenBuffer(
screenshot, self.width, self.height)
# experience memory 에
# S(t), S(t + 1), action, reward 를 저장한다
score = self.ar.getInGameScore(screenshot)
reward = (score - prevscore) / 1000.0
prevscore = score
done = False
reward_st = np.append(reward_st, reward)
self.ddpg.add_experience(
x, observation, action, reward, done)
print "## Add experience (action) (reward) (done)", action, reward, done
# critie, actor network 학습
if counter > TRAIN_STEP:
self.ddpg.train()
reward_per_episode += reward
counter += 1
steps += 1
# 일반적인 상황이 아닌 상황들에 대한 예외처리
elif state == state.LEVEL_SELECTION:
print "unexpected level selection page, go to the last current level: %d" % self.currentLevel
self.ar.loadLevel(self.currentLevel)
elif state == state.MAIN_MENU:
print"unexpected main menu page, reload the level: %d" % self.currentLevel
self.ar.loadLevel(self.currentLevel)
elif state == state.EPISODE_MENU:
print "unexpected episode menu page, reload the level: %d" % self.currentLevel
self.ar.loadLevel(self.currentLevel)
total_reward += reward_per_episode # episode 들의 reward 를 누계
avg_reward = total_reward / episodes
print "## Average reward per episode is : ", avg_reward
def main():
experiment = 'MountainCarContinuous-v0'
env = gym.make(experiment)
steps = env.spec.timestep_limit
assert isinstance(env.observation_space, Box)
assert isinstance(env.action_space, Box)
agent = DDPG(env, is_batch_norm) #这个在循环前面，所以所有的weight都有继承
#也就是说，整个过程只训练了一个模型出来。
exploration_noise = OUNoise(env.action_space.shape[0])
reward_per_episode = 0
total_reward = 0
counter = 0
num_states = env.observation_space.shape[0] - 1
num_actions = env.action_space.shape[0]
#这是state的维度和action的维度
print 'Number of States:', num_states
print 'Number of Actions:', num_actions
print 'Number of steps per episode:', steps
if is_exploration == True:
print("\nExploration phase for {} steps. ".format(exploration_steps))
e_steps = 0
while e_steps < exploration_steps:
s = env.reset()
one_step = 0
done = False
exploration_noise.reset()
exp = []
while not done:
a = exploration_noise.noise()
ss, r, done, _ = env.step(a)
exp.append((s[:-1], a, ss[:-1], r, done))
s = ss
one_step += 1
if one_step > 998:
break
agent.add_experience(exp)
e_steps += 1
reward_st = np.array([0]) #这个是用来存每一次的rewards的
for i in xrange(episodes): #一共要循环1000次
print '====starting episode no:', i, '====', '\n'
observation = env.reset() #每个情节初始化，但是模型参数不初始化
reward_per_episode = 0
LSTM_SIZE = 40
statec_t1 = np.zeros((BATCH_SIZE, LSTM_SIZE))
stateh_t1 = np.zeros((BATCH_SIZE, LSTM_SIZE))
exp = []
for t in xrange(steps):
#env.render()
x = [observation[0:num_states]]
x = np.reshape(x * BATCH_SIZE, [BATCH_SIZE, num_states])
actor, statec_t1, stateh_t1 = agent.evaluate_actor(
x, statec_t1, stateh_t1)
noise = exploration_noise.noise()
#ra = random.random()
if (i < 500):
action = actor[0] + noise
else:
action = actor[0]
observation, reward, done, info = env.step(action)
#print 'Action at step',t,':',action,'reward:',reward,'\n'
exp.append((x, action, observation[0:num_states], reward, done))
if counter > 64:
agent.train()
counter += 1
reward_per_episode += reward
if (done or (t == steps - 1)):
#一个情节结束了～
agent.add_experience(exp)
print 'EPISODE:', i, 'Steps', t, 'Total Reward:', reward_per_episode
print 'Printing reward to file'
exploration_noise.reset()
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, newline='\n')
print '\n\n'
break
total_reward += reward_per_episode
#这里是计算平均值的
print "Average reward per episode {}".format(total_reward / episodes)
def s2l():
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
num_states = feature_size #num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
print("Number of States:", num_states)
print("Number of Actions:", num_actions)
agent = DDPG(env, is_batch_norm, num_states, num_actions)
exploration_noise = OUNoise(env.action_space.shape[0])
counter = 0
total_reward = 0
print("Number of Rollouts per episode:", num_rollouts)
print("Number of Steps per roll out:", steps)
reward_st = np.array([0]) #saving reward
eval_metric_st = np.array([0])
reward_st_all = np.array([0]) #saving reward after every step
frame_obj = Frame_Feature()
#activity_obj=Vid_Feature()
demo_vid_array = demo_array_extractor(demo_folder)
demo_features = frame_obj.video_feature_extractor(demo_vid_array)
for episode in range(num_episodes):
print("==== Starting episode no:", episode, "====", "\n")
env.reset() # Reset env in the begining of each episode
env.render()
obs_img = env.render(mode='rgb_array') # Get the observation
obs_img = np.array(misc.imresize(obs_img, [112, 112, 3]))
observation = np.array(frame_obj.frame_feature_extractor(obs_img))
observation = observation.reshape(-1)
reward_per_episode = 0
for t in range(num_rollouts):
reward_per_rollout = 0
vid_robo_ = []
for i in range(steps):
x = observation
action = agent.evaluate_actor(np.reshape(x, [1, num_states]))
noise = exploration_noise.noise()
action = action[
0] + noise #Select action according to current policy and exploration noise
print('Action at episode-', episode, 'rollout-', t, 'step-', i,
" :", action)
_, _, done, info = env.step(action)
env.render()
obs_robo_ = env.render(mode='rgb_array') # Get the observation
obs_robo = misc.imresize(obs_robo_, [112, 112, 3])
vid_robo_.append(obs_robo)
observation = np.array(
frame_obj.frame_feature_extractor(np.array(obs_robo)))
observation = observation.reshape(-1)
#pasue()
if (i == 15):
vid_robo = np.array(vid_robo_)
robo_features = frame_obj.video_feature_extractor(vid_robo)
reward = -(distance(demo_features, robo_features))
reward = np.array(reward)
print('reward: ', reward)
else:
reward = 0
reward = np.array(reward)
print('reward: ', reward)
# Printing eval_metric after every rollout
eval_metric = np.array(env.get_eval())
eval_metric = eval_metric.reshape(-1)
print('Distance to goal:', eval_metric)
eval_metric_st = np.append(eval_metric_st, eval_metric)
np.savetxt('eval_metric_per_step.txt',
eval_metric_st,
newline="\n")
# Storing reward after every rollout
reward_st_all = np.append(reward_st_all, reward)
np.savetxt('reward_all.txt', reward_st_all, newline="\n")
#add s_t,s_t+1,action,reward to experience memory
agent.add_experience(x, observation, action, reward, False)
reward_per_rollout += reward
counter += 1
#train critic and actor network
if counter > start_training:
agent.train()
print('\n\n')
#Saving policy
if ((episode % 50) == 0 and t == num_rollouts - 1):
print('saving policy...........................!')
agent.save_actor(episode)
reward_per_episode += reward_per_rollout
#check if episode ends:
print('EPISODE: ', episode, ' Total Reward: ', reward_per_episode)
print("Printing reward to file")
exploration_noise.reset(
) #reinitializing random noise for action exploration
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, fmt='%f', newline="\n")
print('\n\n')
total_reward += reward_per_episode
print("Average reward per episode {}".format(total_reward / num_episodes))
def main():
enable_actuator_dynamics = True
env = ControlSystem(enable_actuator_dynamics=enable_actuator_dynamics)
steps = env.timestep_limit #steps per episode
assert isinstance(env.observation_space,
Box), "observation space must be continuous"
assert isinstance(env.action_space, Box), "action space must be continuous"
#Randomly initialize critic,actor,target critic, target actor network and replay buffer
agent = DDPG(env, is_batch_norm)
# agent.load_model()
exploration_noise = OUNoise(env.action_space.shape[0])
counter = 0
reward_per_episode = 0
total_reward = 0
num_states = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
print("Number of States:", num_states)
print("Number of Actions:", num_actions)
print("Number of Steps per episode:", steps)
#saving reward:
reward_st = np.array([0])
log_dir = os.path.join(os.getcwd(), 'logs',
datetime.now().strftime('%Y-%m-%d_%H-%M-%S'),
'action')
if enable_actuator_dynamics == True:
filtered_log_dir = os.path.join(
os.getcwd(), 'logs',
datetime.now().strftime('%Y-%m-%d_%H-%M-%S'), 'filtered_action')
y_hat_log_dir = os.path.join(os.getcwd(), 'logs',
datetime.now().strftime('%Y-%m-%d_%H-%M-%S'),
'y_hat')
y_ref_log_dir = os.path.join(os.getcwd(), 'logs',
datetime.now().strftime('%Y-%m-%d_%H-%M-%S'),
'y_ref')
gen_function_log_dir = os.path.join(
os.getcwd(), 'logs',
datetime.now().strftime('%Y-%m-%d_%H-%M-%S'), 'function')
os.makedirs(log_dir)
if enable_actuator_dynamics == True:
os.makedirs(filtered_log_dir)
os.makedirs(y_hat_log_dir)
os.makedirs(y_ref_log_dir)
os.makedirs(gen_function_log_dir)
for i in range(episodes):
print("==== Starting episode no:", i, "====")
observation = env.reset()
reward_per_episode = 0
actions_per_episode = []
if enable_actuator_dynamics == True:
filtered_action_per_episode = []
for t in range(steps):
#rendering environmet (optional)
env.render()
x = observation
action = agent.evaluate_actor(np.reshape(x, [1, num_states]))
noise = exploration_noise.noise()
action = action[
0] + noise #Select action according to current policy and exploration noise
actions_per_episode.append(action)
# if i % 100 == 0:
# print ("Action at step", t ," :",action,"\n")
# print("#", action[0])
if action[0] < 0:
action = [0]
elif action[0] > 1:
action = [1]
# print("Step", t, 'action', action)
if enable_actuator_dynamics == False:
observation, reward, Y_plot, t_plot, y_ref, random_function = env.step(
action, t)
elif enable_actuator_dynamics == True:
observation, reward, filtered_action, Y_plot, t_plot, y_ref, random_function = env.step(
action, t)
filtered_action_per_episode.append(filtered_action)
# print ("Reward at step", t ," :",reward,"\n")
#add y_t,y_t-1,action,reward,timestep to experience memory
agent.add_experience(x, observation, action, reward, t)
#train critic and actor network
if counter > 64:
agent.train()
reward_per_episode += reward
counter += 1
#check if episode ends:
if (t == steps - 1):
print('EPISODE: ', i, ' Steps: ', t, ' Total Reward: ',
reward_per_episode)
# print ("Printing reward to file")
exploration_noise.reset(
) #reinitializing random noise for action exploration
reward_st = np.append(reward_st, reward_per_episode)
np.savetxt('episode_reward.txt', reward_st, newline="\n")
# print("Y_plot")
# plt.step(t_plot,Y_plot)
# plt.grid()
# plt.xlabel('t')
# plt.ylabel('y')
# plt.show()
# Save actions
np.savetxt(log_dir + '/' + str(i).zfill(7) + '.txt',
actions_per_episode)
if enable_actuator_dynamics == True:
np.savetxt(
filtered_log_dir + '/' + str(i).zfill(7) + '.txt',
filtered_action_per_episode)
np.savetxt(y_hat_log_dir + '/' + str(i).zfill(7) + '.txt',
Y_plot)
np.savetxt(y_ref_log_dir + '/' + str(i).zfill(7) + '.txt',
y_ref)
# np.savetxt(gen_function_log_dir + '/' + str(i).zfill(7) + '.txt', random_function)
# save model
if i % 100 == 0:
print('save')
agent.save_model()
# print ('\n\n')
break
total_reward += reward_per_episode
print("Average reward per episode {}".format(total_reward / episodes))