track_number=track_number)
agent = Agent(env)
print("TORCS Experiment Start.")
for i in range(episode_count):
print("Episode : " + str(i))
if np.mod(i, 3) == 0:
# Sometimes you need to relaunch TORCS because of the memory leak error
ob = env.reset(relaunch=True)
else:
ob = env.reset()
total_reward = 0.
for j in range(max_steps):
action = agent.act(ob)
ob, reward, done, _ = env.step(action)
# print(ob)
total_reward += reward
step += 1
if done:
break
print("TOTAL REWARD @ " + str(i) + " -th Episode : " + str(total_reward))
print("Total Step: " + str(step))
print("")
env.end() # This is for shutting down TORCS
print("Finish.")
print("TORCS Experiment Start.")
for i in range(episode_count):
print("Episode : " + str(i))
if np.mod(i, 3) == 0:
# Sometimes you need to relaunch TORCS because of the memory leak error
ob = env.reset(relaunch=True)
else:
ob = env.reset()
# env.reset() return current state including
#
total_reward = 0.
for j in range(max_steps):
action = agent.act(ob, reward, done, vision)
ob, reward, done, _ = env.step(action)
total_reward += reward
step += 1
if done:
break
print("TOTAL REWARD @ " + str(i) + " -th Episode : " + str(total_reward))
print("Total Step: " + str(step))
print("")
env.end() # This is for shutting down TORCS
print("Finish.")
agent = Agent(1) # steering only
print("TORCS Experiment Start.")
for i in range(episode_count):
print("Episode : " + str(i))
if np.mod(i, 3) == 0:
# Sometimes you need to relaunch TORCS because of the memory leak error
ob = env.reset(relaunch=True)
else:
ob = env.reset()
total_reward = 0.
for j in range(max_steps):
action = agent.act(ob, reward, done, vision)
ob, reward, done, _ = env.step(action)
#print(ob)
total_reward += reward
step += 1
if done:
break
print("TOTAL REWARD @ " + str(i) +" -th Episode : " + str(total_reward))
print("Total Step: " + str(step))
print("")
env.end() # This is for shutting down TORCS
print("Finish.")
if __name__ == "__main__":
log_level = logging.INFO
logging.basicConfig(format='%(asctime)s %(levelname)-8s %(message)s',
level=log_level,
datefmt='%Y-%m-%d %H:%M:%S')
logging.info("simstar env init")
env = SimstarEnv(synronized_mode=True, speed_up=5, hz=1)
agent = Agent(dim_action=3)
logging.info("entering main loop")
for ee in range(max_episodes):
episode_total_reward = 0
logging.info("reset environment. eposde number: %d", ee + 1)
time.sleep(2)
observation = env.reset()
for ii in range(max_steps):
action = agent.act(observation, reward, done)
action[2] = 0.0
print(ii, action)
debug_action = [0.0, 1.0, 0.0]
action = debug_action
observation, reward, done, _ = env.step(action)
episode_total_reward += reward
#time.sleep(0.5)
if (done):
break
env.clear()
ob = env.reset()
total_reward = 0.
states = np.array([[0, 0, 0]])
partial_delta_theta = np.array([np.zeros((n_action,n_states))])
J = 0
step = 0
delta_theta = 0
baseline_n = np.array([np.zeros((n_action,n_states))])
baseline_d = np.array([np.zeros((n_action,n_states))])
J_vector = np.array([[0]])
for step in range(max_steps):
action, av_theta, ob_theta = agent.act(ob, reward, done, vision, theta)
ob, reward, done, _ = env.step(action)
#print("\n-------------------------------------------------------")
#print(ob)
#print("\n-------------------------------------------------------")
total_reward += reward
## update the vector of trajectories
states = np.append(states, [[ob_theta, action, av_theta]], axis=0 )
## update performance
J= J + (0.99**step) * (reward)
#Compute delta_theta till the current partial sum
delta_theta = delta_theta + compute_Delta_Theta(states)
