def run(self):
env = gym.make("CartPole-v0") # Create the environment
episode_done = True
trajectory = Trajectory()
next_state = []
episode_num = 0
episode_reward = 1 # TODO: check for alternative init rewards
while True:
if self.stop: # Check if we need to stop the current thread
print("Worker:",self.worker_id, "STOP")
break
if episode_done: # Check if the episode is done
state = env.reset() # Reset the environment and recover initial state
episode_done = False
template = "in worker {}, episode {} done after {} steps" # Template for print
print(template.format(self.worker_id, episode_num, episode_reward))
self.res_queue.put(episode_reward)
episode_num += 1
episode_reward = 1
for i in range(LOOKAHEAD):
#print("I=",i)
if self.is_eval(): # Render curretn state of the environement
env.render()
state = tf.convert_to_tensor(state) # Convert state to tensor
state = tf.expand_dims(state, 0)
action = self.act(state) # Select action based on the current policy.
trajectory.store(s=state, a=action)
state, reward, episode_done, _ = env.step(action) # Recover next state, reward and if the episode is done.
#print("|New state\n",state,"\n|reward: \n",reward," \n|episode done \n",episode_done)
if episode_done: # If the agent fails the task
reward = -1
# Store the reward
trajectory.store(r=reward)
episode_reward += reward
next_state = state # update current state
if episode_done:
next_state = []
break
if episode_done and self.is_eval():
if self.is_eval():
env.close()
template = "in worker {}, episode {} done after {} steps" # Template for print
print(template.format(self.worker_id, episode_num, episode_reward))
break
if not self.is_eval():
self.train(trajectory, next_state)# Update network using the trajectory
trajectory.clear()
def run(self):
env = gym.make("CartPole-v0")
episode_done = True
trajectory = Trajectory()
next_state = []
episode_num = 0
episode_reward = 1
while True:
if episode_done:
state = env.reset()
episode_done = False
template = "in worker {}, episode {} done after {} steps"
print(
template.format(self.worker_id, episode_num,
episode_reward))
self.res_queue.put(episode_reward)
episode_num += 1
episode_reward = 1
for i in range(LOOKAHEAD):
# env.render()
state = tf.convert_to_tensor(state)
state = tf.expand_dims(state, 0)
action = self.act(state)
trajectory.store(s=state, a=action)
state, reward, episode_done, _ = env.step(action)
if episode_done:
reward = -1
trajectory.store(r=reward)
episode_reward += reward
next_state = state
if episode_done:
next_state = []
break
self.train(trajectory, next_state)
trajectory.clear()
