def Monte_Carlo(self):
"""
Monte_Carlo experiments
:return: Q_table, policy
"""
Q_table = self.table_init() # Q_table initialization
policy = {} # policy table
for k in range(self.M): # iterations
x = self.state_init() # initial state
u = self.epsilon_greedy(int(np.argmax(Q_table[x])), self.epsilon)
while x != self.xG: # stop condition
x_next = self.move_next(x, self.u_set[u]) # next state
reward = env.get_reward(x_next, self.lose) # reward observed
u_next = self.epsilon_greedy(int(np.argmax(Q_table[x_next])),
self.epsilon)
Q_table[x][u] = (1 - self.alpha) * Q_table[x][u] + \
self.alpha * (reward + self.gamma * Q_table[x_next][u_next])
x, u = x_next, u_next
for x in Q_table:
policy[x] = int(np.argmax(Q_table[x])) # extract policy
return Q_table, policy
def cal_Q_value(self, x, p, table):
"""
cal Q_value.
:param x: next state vector
:param p: probability of each state
:param table: value table
:return: Q-value
"""
value = 0
reward = env.get_reward(x, self.xG,
self.lose) # get reward of next state
for i in range(len(x)):
value += p[i] * (reward[i] + self.gamma * max(table[x[i]]))
return value
env.dump_image(os.path.join(img_dir, '%d.png' % t))
view_batches = get_view(env) # s
actions, actions_batches = model.infer_actions(
sess, view_batches, policy=argv.policy,
epsilon=argv.epsilon) # a
env.take_action(actions)
env.decrease_health()
env.update_pig_pos()
# env.update_rabbit_pos()
if video_flag:
env.dump_image(os.path.join(img_dir, '%d.png' % (t + 1)))
rewards = get_reward(env) # r, a dictionary
env.increase_health(rewards)
total_reward = 0
for k, v in rewards.items():
total_reward += v
new_view_batches = get_view(env) # s'
maxQ_batches = model.infer_max_action_values(
sess, new_view_batches)
model.train(sess=sess,
view_batches=view_batches,
actions_batches=actions_batches,
rewards=rewards,
maxQ_batches=maxQ_batches,
learning_rate=argv.learning_rate)
predict = main_qn.predict(predict_arr)[0]
action = np.argmax(predict)
value = predict
do_action(action)
time.sleep(0.05)
state = get_state()
state_deque.append(state.astype('float32') / 255.0)
print('Episode: {}, Step: {}, Epsilon: {}, Action: {}, '.format(
episode, step, epsilon, action_name[action]),
end='')
if step > warmup_steps:
reward = get_reward(state_deque)
if is_scrolling(state_deque) and (action == 2 or action == 3):
reward += 0.1 # additional reward for moving toward the right side
if is_dead(state_deque):
reward = -1
dead = True
print('Reward: {}'.format(reward))
memory.add((current_state_arr, action, reward,
get_state_arr(state_deque)))
print(value)
if do_learn is True:
if len(memory) >= batch_size:
pause_button()
0, top_action.data[0]] + prob_choose_top
sampled_action = Categorical(action_probs).sample()
print(action_probs)
print(epsilon)
# Act and update the current observation
cur_observation, cur_frame, done = env.act(agent_host,
sampled_action.data[0])
num_timesteps += 1
if (done):
break
# Calculate the reward based on the current task
reward = env.get_reward(task, cur_observation, prev_observation)
# Check for episode's termination
if (reward > 0):
done = True
if (num_timesteps >= max_EPISODE_LENGTH):
done = True
episode_trajectory.append(
Point(last_frames.numpy(), instruction.numpy(),
sampled_action.data.numpy(), reward,
action_probs.data.numpy()))
if (done):
break
def apply_action_in_env(states, actions):
lats = env.convert_to_latents(states)
new_lats = env.apply_actions(lats, actions, simplify=False)
rewards = env.get_reward(lats, actions, new_lats, simplify=False)
new_states = env.convert_to_obs(lats)
return [rewards, new_states]
# function and observe the reward value. We then update the record array with this new
# observation. We repeat this process a bunch of times, and it will continually update
# the record array. The arm with the highest reward probability should eventually get
# chosen most often, since it will give out the highest average reward.
# excerpt from the Book: Deep Reinforcement Learning in Action A.Zai, B. Brown
for i in range(500):
if (use_softmax):
# use softmax probabilities of values to choose action
p = softmax(record[:, 1])
choice = np.random.choice(np.arange(n), p=p)
else:
# use for epsilon greedy approach
if random.random() > eps:
choice = get_best_arm(record)
else:
choice = np.random.randint(10)
# computes the reward for choosing the arm
r = get_reward(probs[choice])
# updates the Q-tale
record = update_record(record, choice, r)
#print(f"Updated record: {record}")
#keeps track of the running average of rewards
mean_reward = ((i + 1) * rewards[-1] + r) / (i + 2)
rewards.append(mean_reward)
# plot the cumalative rewards vs iteration
ax.scatter(np.arange(len(rewards)), rewards)
plt.savefig("./cumulative_rewards.png")