def mountain_car(self, i):
qrl = MountainCar(env_name='MountainCar-v0',
learning_rate=0.97**i,
discount=0.99**i,
iterations=1000)
qrl.run()
num_rewards = len(qrl.rewards)
return self.rolling_mean(qrl.rewards,
window=num_rewards // 100,
strides=num_rewards // 50)
def fig_10_2():
fig, ax = plt.subplots()
qhat, nab_qhat = get_qhats(N_TIL, N_TLGS)
for alp in FIG_10_2_ALP_L:
print(f"[alpha={alp}]")
tot_n_steps = np.zeros(FIG_10_2_N_EP)
for seed in range(FIG_10_2_N_RUNS):
print(f"[RUN #{seed}]")
alg = EpisodicSemiGradientTD0(MountainCar(),
alp,
N_TIL * N_TLGS,
eps=0)
alg.seed(seed)
tot_n_steps += np.array(
alg.pol_eva(qhat, nab_qhat, FIG_10_2_N_EP, FIG_10_2_G))
plt.plot(tot_n_steps / FIG_10_2_N_RUNS, label=f'alpha={alp}')
plt.yscale('log')
xticks, yticks = [0, 500], [100, 200, 400, 1000]
plot_figure(ax, 'Figure 10.2', xticks, xticks, 'Episode', yticks, yticks,
'Steps\nper episode\n(log scale)')
fig.set_size_inches(20, 14)
plt.legend()
save_plot('fig10.2', dpi=100)
plt.show()
def fig_10_4():
fig, ax = plt.subplots()
qhat, nab_qhat = get_qhats(N_TIL, N_TLGS)
alg = nStepSemiGradSarsa(MountainCar(), 0, N_TIL * N_TLGS, 0, 0)
for n in FIG_10_4_ALP_BND:
alg.n = n
print(f"[n={n}]")
steps_l = []
alpha_l = np.linspace(*FIG_10_4_ALP_BND[n], FIG_10_4_ALP_PTS)
for alpha in alpha_l:
alg.a = alpha / N_TLGS
print(f"[alpha={alg.a}]")
tot_steps = 0
for seed in range(FIG_10_4_N_RUNS):
print(f"[RUN #{seed}]")
alg.reset()
alg.seed(seed)
for ep in range(FIG_10_4_N_EP):
tot_steps += alg.pol_eva(None,
qhat,
nab_qhat,
1,
FIG_10_4_G,
max_steps=1000)[0]
steps_l.append(tot_steps / (FIG_10_4_N_RUNS * FIG_10_4_N_EP))
plt.plot(alpha_l, steps_l, label=f'n={n}')
xticks, yticks = np.linspace(0, 1.5, 4), np.linspace(220, 300, 5)
left_title = (
f'Mountain Car\nSteps per\nepisode\n(averaged \nover ' +
f'first\n{FIG_10_4_N_EP} episodes\nand {FIG_10_4_N_RUNS} runs')
plot_figure(ax,
'Figure 10.4',
list(xticks) + [1.8],
xticks,
f'alpha * number of tilings ({N_TLGS})',
yticks,
yticks,
left_title,
labelpad=20)
fig.set_size_inches(20, 14)
plt.legend()
save_plot('fig10.4', dpi=100)
plt.show()
def fig_12_11():
fig, ax = plt.subplots()
F, qhat = get_fn_mc(N_TIL, N_TLGS)
for alg_name in FIG_12_11_ALG_STR.keys():
steps_l = []
alpha_l = np.linspace(*FIG_12_11_ALP_BND[alg_name], FIG_12_11_N_PTS)
for alpha in alpha_l:
alg = alg_name(MountainCar(), alpha / N_TLGS, N_TIL * N_TLGS,
FIG_12_11_LAM, F, qhat, FIG_12_11_EPS, FIG_12_11_G)
print(f"[ALPHA={alg.a}]")
tot_steps = 0
for seed in range(FIG_12_11_N_RUNS):
print(f"[RUN #{seed}]")
alg.reset()
alg.seed(seed)
for ep in range(FIG_12_11_N_EP):
tot_steps += alg.pol_eva(None,
1,
max_steps=FIG_12_11_MAX_STEPS)[0]
steps_l.append(tot_steps / (FIG_12_11_N_RUNS * FIG_12_11_N_EP))
plt.plot(alpha_l,
-np.array(steps_l),
label=FIG_12_11_ALG_STR[alg_name])
xticks, yticks = np.linspace(0.2, 2, 10), np.linspace(-550, -150, 9)
xnames = map(lambda x: str(x)[:3], xticks)
left_title = (
f'Mountain Car\nReward per\nepisode\n(averaged \nover ' +
f'first\n{FIG_12_11_N_EP} episodes\n{FIG_12_11_N_RUNS} runs)')
plot_figure(ax,
'Figure 12.11',
xticks,
xnames,
f'alpha * number of tilings ({N_TLGS})',
yticks,
yticks,
left_title,
labelpad=45)
fig.set_size_inches(20, 14)
plt.legend()
save_plot('fig12.11', dpi=100)
plt.show()
def fig_12_10():
fig, ax = plt.subplots()
for lam in FIG_12_10_LAM_L:
print(f"[LAM={lam}]")
steps_l = []
alpha_l = np.linspace(FIG_12_10_ALP_MIN, FIG_12_10_ALP_MAX,
FIG_12_10_N_PTS)
for alpha in alpha_l:
F, qhat = get_fn_mc(N_TIL, N_TLGS)
alg = SarsaLam(MountainCar(), alpha / N_TLGS, N_TIL * N_TLGS, lam,
F, qhat, FIG_12_10_EPS, FIG_12_10_G)
print(f"[ALPHA={alg.a}]")
tot_steps = 0
for seed in range(FIG_12_10_N_RUNS):
print(f"[RUN #{seed}]")
alg.reset()
alg.seed(seed)
for ep in range(FIG_12_10_N_EP):
print(f"[EP #{ep}]")
tot_steps += alg.pol_eva(None,
1,
max_steps=FIG_12_10_MAX_STEPS)[0]
steps_l.append(tot_steps / (FIG_12_10_N_RUNS * FIG_12_10_N_EP))
plt.plot(alpha_l, steps_l, label=f'lam={lam}')
xticks, yticks = np.linspace(0.5, 1.5, 5), np.linspace(180, 300, 7)
left_title = (
f'Mountain Car\nSteps per\nepisode\n(averaged \nover ' +
f'first\n{FIG_12_10_N_EP} episodes\n{FIG_12_10_N_RUNS} runs)')
plot_figure(ax,
'Figure 12.10',
list(xticks) + [1.6],
xticks,
f'alpha * number of tilings ({N_TLGS})', [160] + list(yticks),
yticks,
left_title,
labelpad=35)
fig.set_size_inches(20, 14)
plt.legend()
save_plot('fig12.10', dpi=100)
plt.show()
def fig_10_3():
fig, ax = plt.subplots()
qhat, nab_qhat = get_qhats(N_TIL, N_TLGS)
for (n, alp) in zip(FIG_10_3_N_L, FIG_10_3_ALP_L):
print(f"[n={n}, alpha={alp}]")
tot_n_steps = np.zeros(FIG_10_3_N_EP)
for seed in range(FIG_10_3_N_RUNS):
print(f"[RUN #{seed}]")
alg = nStepSemiGradSarsa(MountainCar(), alp, N_TIL * N_TLGS, 0, n)
alg.seed(seed)
tot_n_steps += np.array(
alg.pol_eva(None, qhat, nab_qhat, FIG_10_3_N_EP, FIG_10_3_G))
plt.plot(tot_n_steps / FIG_10_3_N_RUNS, label=f'n={n}')
plt.yscale('log')
xticks, yticks = [0, 500], [100, 200, 400, 1000]
plot_figure(ax, 'Figure 10.3', xticks, xticks, 'Episode', yticks, yticks,
'Steps\nper episode\n(log scale)')
fig.set_size_inches(20, 14)
plt.legend()
save_plot('fig10.3', dpi=100)
plt.show()
def fig_10_1():
def plot_and_save(filename, title, alg, n_ep, max_steps=np.inf):
fig = plt.figure()
alg.pol_eva(qhat, nab_qhat, n_ep, FIG_10_1_G, max_steps=max_steps)
print_qhat_mc(alg, fig, '111', title)
fig.set_size_inches(20, 14)
save_plot(filename, dpi=100)
plt.show()
qhat, nab_qhat = get_qhats(N_TIL, N_TLGS)
env = MountainCar()
alg = EpisodicSemiGradientTD0(env, FIG_10_1_ALP, N_TIL * N_TLGS, eps=0)
alg.seed(0)
plot_and_save(f'fig10.1_{FIG_10_1_STEPS}_steps', f'Step {FIG_10_1_STEPS}',
alg, 1, FIG_10_1_STEPS)
tot_ep = 1
for ep in FIG_10_1_EP_L:
alg.pol_eva(qhat, nab_qhat, ep - tot_ep, FIG_10_2_G)
plot_and_save(f'fig10.1_{ep}_episodes', f'Episode {ep}', alg,
ep - tot_ep)
tot_ep += (ep - tot_ep)
def main():
env = MountainCar(mass=0.2, friction=0.3, delta_t=0.1)
# Define the state arrays for velocity and position
tot_action = 3 # Three possible actions
tot_bins = 12 # the value used to discretize the space
velocity_state_array = np.linspace(-1.5,
+1.5,
num=tot_bins - 1,
endpoint=False)
position_state_array = np.linspace(-1.2,
+0.5,
num=tot_bins - 1,
endpoint=False)
# Random policy as a square matrix of size (tot_bins x tot_bins)
# Three possible actions represented by three integers
policy_matrix = np.random.randint(low=0,
high=tot_action,
size=(tot_bins,
tot_bins)).astype(np.float32)
print("Policy Matrix:")
print(policy_matrix)
# The state-action matrix and the visit counter
# The rows are the velocities and the columns the positions.
state_action_matrix = np.zeros((tot_action, tot_bins * tot_bins))
visit_counter_matrix = np.zeros((tot_action, tot_bins * tot_bins))
# Variables
gamma = 0.999
alpha = 0.001
tot_episode = 100000
epsilon_start = 0.9 # those are the values for epsilon decay
epsilon_stop = 0.1
epsilon_decay_step = 3000
print_episode = 500 # print every...
movie_episode = 10000 # movie saved every...
reward_list = list()
step_list = list()
for episode in range(tot_episode):
epsilon = return_decayed_value(epsilon_start,
epsilon_stop,
episode,
decay_step=epsilon_decay_step)
# Reset and return the first observation
observation = env.reset(exploring_starts=False)
# The observation is digitized, meaning that an integer corresponding
# to the bin where the raw float belongs is obtained and use as replacement.
observation = (np.digitize(observation[1], velocity_state_array),
np.digitize(observation[0], position_state_array))
is_starting = True
cumulated_reward = 0
for step in range(100):
#Take the action from the action matrix
#action = policy_matrix[observation[0], observation[1]]
#Take the action using epsilon-greedy
action = return_epsilon_greedy_action(policy_matrix,
observation,
epsilon=epsilon)
if (is_starting):
action = np.random.randint(0, tot_action)
is_starting = False
#Move one step in the environment and get obs and reward
new_observation, reward, done = env.step(action)
new_observation = (np.digitize(new_observation[1],
velocity_state_array),
np.digitize(new_observation[0],
position_state_array))
new_action = policy_matrix[new_observation[0], new_observation[1]]
#Updating the state-action matrix
state_action_matrix = update_state_action(
state_action_matrix, visit_counter_matrix, observation,
new_observation, action, new_action, reward, alpha, gamma)
#Updating the policy
policy_matrix = update_policy(policy_matrix, state_action_matrix,
observation)
#Increment the visit counter
visit_counter_matrix = update_visit_counter(
visit_counter_matrix, observation, action)
observation = new_observation
cumulated_reward += reward
if done: break
# Store the data for statistics
reward_list.append(cumulated_reward)
step_list.append(step)
# Printing utilities
if (episode % print_episode == 0):
print("")
print("Episode: " + str(episode + 1))
print("Epsilon: " + str(epsilon))
print("Episode steps: " + str(step + 1))
print("Cumulated Reward: " + str(cumulated_reward))
print("Policy matrix: ")
print_policy(policy_matrix)
if (episode % movie_episode == 0):
print("Saving the reward plot in: ./reward.png")
plot_curve(reward_list,
filepath="./reward.png",
x_label="Episode",
y_label="Reward",
x_range=(0, len(reward_list)),
y_range=(-1.1, 1.1),
color="red",
kernel_size=500,
alpha=0.4,
grid=True)
print("Saving the step plot in: ./step.png")
plot_curve(step_list,
filepath="./step.png",
x_label="Episode",
y_label="Steps",
x_range=(0, len(step_list)),
y_range=(-0.1, 100),
color="blue",
kernel_size=500,
alpha=0.4,
grid=True)
print("Saving the gif in: ./mountain_car.gif")
env.render(file_path='./mountain_car.gif', mode='gif')
print("Complete!")
# Save reward and steps in npz file for later use
# np.savez("./statistics.npz", reward=np.asarray(reward_list), step=np.asarray(step_list))
# Time to check the utility matrix obtained
print("Policy matrix after " + str(tot_episode) + " episodes:")
print_policy(policy_matrix)
def main():
env = MountainCar(mass=0.2, friction=0.3, delta_t=0.1)
# Define the state arrays for velocity and position
tot_action = 3 # Three possible actions
tot_bins = 12 # the value used to discretize the space
velocity_state_array = np.linspace(-1.5, +1.5, num=tot_bins-1,
endpoint=False)
position_state_array = np.linspace(-1.2, +0.5, num=tot_bins-1,
endpoint=False)
# Random policy as a square matrix of size (tot_bins x tot_bins)
# Three possible actions represented by three integers
policy_matrix = np.random.randint(low=0, high=tot_action, size=(tot_bins,tot_bins))
print("Policy Matrix:")
print(policy_matrix)
# The state-action matrix and the visit counter
# The rows are the velocities and the columns the positions.
state_action_matrix = np.zeros((tot_action, tot_bins*tot_bins))
visit_counter_matrix = np.zeros((tot_action, tot_bins*tot_bins))
# Variables
gamma = 0.999
alpha = 0.001
tot_episode = 100000
epsilon_start = 0.9 # those are the values for epsilon decay
epsilon_stop = 0.1
epsilon_decay_step = 3000
print_episode = 500 # print every...
movie_episode = 10000 # movie saved every...
reward_list = list()
step_list = list()
for episode in range(tot_episode):
epsilon = return_decayed_value(epsilon_start, epsilon_stop, episode,
decay_step=epsilon_decay_step)
# Reset and return the first observation
observation = env.reset(exploring_starts=False)
# The observation is digitized, meaning that an integer corresponding
# to the bin where the raw float belongs is obtained and use as replacement.
observation = (np.digitize(observation[1], velocity_state_array),
np.digitize(observation[0], position_state_array))
is_starting = True
cumulated_reward = 0
for step in range(100):
#Take the action from the action matrix
#action = policy_matrix[observation[0], observation[1]]
#Take the action using epsilon-greedy
action = return_epsilon_greedy_action(policy_matrix, observation,
epsilon=epsilon)
if(is_starting):
action = np.random.randint(0, tot_action)
is_starting = False
#Move one step in the environment and get obs and reward
new_observation, reward, done = env.step(action)
new_observation = (np.digitize(new_observation[1],
velocity_state_array),
np.digitize(new_observation[0],
position_state_array))
new_action = policy_matrix[new_observation[0], new_observation[1]]
#Updating the state-action matrix
state_action_matrix = update_state_action(state_action_matrix,
visit_counter_matrix,
observation,
new_observation,
action,
new_action,
reward,
alpha,
gamma,
tot_bins)
#Updating the policy
policy_matrix = update_policy(policy_matrix,
state_action_matrix,
observation,
tot_bins)
#Increment the visit counter
visit_counter_matrix = update_visit_counter(visit_counter_matrix,
observation,
action,
tot_bins)
observation = new_observation
cumulated_reward += reward
if done: break
# Store the data for statistics
reward_list.append(cumulated_reward)
step_list.append(step)
# Printing utilities
if(episode % print_episode == 0):
print("")
print("Episode: " + str(episode+1))
print("Epsilon: " + str(epsilon))
print("Episode steps: " + str(step+1))
print("Cumulated Reward: " + str(cumulated_reward))
print("Policy matrix: ")
print_policy(policy_matrix)
if(episode % movie_episode == 0):
print("Saving the reward plot in: ./reward.png")
plot_curve(reward_list, filepath="./reward.png",
x_label="Episode", y_label="Reward",
x_range=(0, len(reward_list)), y_range=(-1.1,1.1),
color="red", kernel_size=500,
alpha=0.4, grid=True)
print("Saving the step plot in: ./step.png")
plot_curve(step_list, filepath="./step.png",
x_label="Episode", y_label="Steps",
x_range=(0, len(step_list)), y_range=(-0.1,100),
color="blue", kernel_size=500,
alpha=0.4, grid=True)
print("Saving the gif in: ./mountain_car.gif")
env.render(file_path='./mountain_car.gif', mode='gif')
print("Complete!")
# Save reward and steps in npz file for later use
# np.savez("./statistics.npz", reward=np.asarray(reward_list), step=np.asarray(step_list))
# Time to check the utility matrix obtained
print("Policy matrix after " + str(tot_episode) + " episodes:")
print_policy(policy_matrix)
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from mountain_car import MountainCar
import random
my_car = MountainCar(mass=0.2, friction=0.3, delta_t=0.1)
cumulated_reward = 0
print("Starting random agent...")
for step in range(100):
action = random.randint(a=0, b=2)
observation, reward, done = my_car.step(action)
cumulated_reward += reward
if done: break
print("Finished after: " + str(step+1) + " steps")
print("Cumulated Reward: " + str(cumulated_reward))
print("Saving the gif in: ./mountain_car.gif")
my_car.render(file_path='./mountain_car.gif', mode='gif')
print("Complete!")
DECAY = 0.99
ALPHA = 0.9
BIN_NUM = 64
train_times = 31250
vis_times = 10
output_path = 'output/mountain_car_sarsa/'
if not os.path.exists(output_path):
os.makedirs(output_path)
def random_action(_, n):
return np.random.randint(n)
bin_shape = (BIN_NUM, BIN_NUM)
games = MountainCar(PARA_SIZE, REPLAY_SIZE)
inc_learner = BinIncLearner(Discretization(state_bounds, bin_shape), bin_shape,
(action_n, ), ALPHA)
framework = SARSA(games, inc_learner, random_action, DECAY)
for v in xrange(vis_times):
framework.loop(train_times)
visualize_states(output_path + 'states_%i.png' % v,
inc_learner.eval_batch_no_default)
games.close()
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from mountain_car import MountainCar
import random
my_car = MountainCar(mass=0.2, friction=0.3, delta_t=0.1)
cumulated_reward = 0
print("Starting random agent...")
for step in range(100):
action = random.randint(a=0, b=2)
observation, reward, done = my_car.step(action)
cumulated_reward += reward
if done: break
print("Finished after: " + str(step + 1) + " steps")
print("Cumulated Reward: " + str(cumulated_reward))
print("Saving the gif in: ./mountain_car.gif")
my_car.render(file_path='./mountain_car.gif', mode='gif')
print("Complete!")
import argparse
import os
from mountain_car import MountainCar
import matplotlib.pyplot as plt
ENV_DICT = {
'mountain_car': MountainCar(mnt=False),
}
def play(env):
def refresh():
os.system('cls' if os.name == 'nt' else 'clear')
print(env)
while True:
env.reset()
done = False
v = []
while not done:
key = ''
while key not in env.keys:
refresh()
key = input("press key\n$>")
if key == "exit()":
exit()
if (key == 'p'):
env.show(n_pts=10000)
if (key == 'v'):
plt.plot(v)
plt.show()
"""The world's simplest agent!"""
def __init__(self, action_space):
self.action_space = action_space
def act(self, observation, reward, done):
return self.action_space.sample()
if __name__ == '__main__':
logging.basicConfig()
log = logging.getLogger("mountain-car")
log.setLevel(level='INFO')
# we will use our environment (wrapper of OpenAI env)
mountain_car = MountainCar()
# specify which agent you want to use,
# BonsaiAgent that uses trained Brain or
# RandomAgent that randomly selects next action
agent = BonsaiAgent()
episode_count = 100
try:
for i in range(episode_count):
#start a new episode and get the new state
mountain_car.episode_start()
state = mountain_car.get_state()
while True: