def experiment(n_epochs, n_episodes):
np.random.seed()
# MDP
n_steps = 5000
mdp = InvertedPendulum(horizon=n_steps)
# Agent
n_tilings = 10
alpha_theta = Parameter(5e-3 / n_tilings)
alpha_omega = Parameter(0.5 / n_tilings)
alpha_v = Parameter(0.5 / n_tilings)
tilings = Tiles.generate(n_tilings, [10, 10],
mdp.info.observation_space.low,
mdp.info.observation_space.high + 1e-3)
phi = Features(tilings=tilings)
input_shape = (phi.size,)
mu = Regressor(LinearApproximator, input_shape=input_shape,
output_shape=mdp.info.action_space.shape)
sigma = 1e-1 * np.eye(1)
policy = GaussianPolicy(mu, sigma)
agent = COPDAC_Q(policy, mu, mdp.info,
alpha_theta, alpha_omega, alpha_v,
value_function_features=phi,
policy_features=phi)
# Train
dataset_callback = CollectDataset()
visualization_callback = Display(agent._V, mu,
mdp.info.observation_space.low,
mdp.info.observation_space.high,
phi, phi)
core = Core(agent, mdp, callbacks=[dataset_callback])
for i in range(n_epochs):
core.learn(n_episodes=n_episodes,
n_steps_per_fit=1, render=False)
J = compute_J(dataset_callback.get(), gamma=1.0)
dataset_callback.clean()
visualization_callback()
print('Mean Reward at iteration ' + str(i) + ': ' +
str(np.sum(J) / n_steps / n_episodes))
print('Press a button to visualize the pendulum...')
input()
sigma = 1e-8 * np.eye(1)
policy.set_sigma(sigma)
core.evaluate(n_steps=n_steps, render=True)
def experiment(mdp, agent_high, agent_low, n_epochs, n_episodes, ep_per_eval,
ep_per_fit_low, display, print_j, quiet):
np.random.seed()
dataset_callback = CollectDataset()
computational_graph = build_computational_graph(mdp, agent_low, agent_high,
ep_per_fit_low,
[dataset_callback])
core = HierarchicalCore(computational_graph)
J_list = list()
L_list = list()
dataset = core.evaluate(n_episodes=ep_per_eval, quiet=quiet)
J = compute_J(dataset, gamma=mdp.info.gamma)
J_list.append(np.mean(J))
J_low_list = list()
L = episodes_length(dataset)
L_list.append(np.mean(L))
if print_j:
print('Reward at start :', J_list[-1])
for n in range(n_epochs):
core.learn(n_episodes=n_episodes, skip=True, quiet=quiet)
ll_dataset = dataset_callback.get()
dataset_callback.clean()
J_low = compute_J(ll_dataset, mdp.info.gamma)
J_low_list.append(np.mean(J_low))
if print_j:
print('Low level reward at epoch', n, ':', np.mean(J_low))
dataset = core.evaluate(n_episodes=ep_per_eval, quiet=quiet)
J = compute_J(dataset, gamma=mdp.info.gamma)
J_list.append(np.mean(J))
L = episodes_length(dataset)
L_list.append(np.mean(L))
if print_j:
print('Reward at epoch ', n, ':', J_list[-1])
if display:
core.evaluate(n_episodes=1, render=True)
return J_list, L_list, J_low_list
def experiment(alg, params, n_epochs, n_episodes, n_ep_per_fit):
np.random.seed()
# MDP
mdp = Segway()
# Policy
approximator = Regressor(LinearApproximator,
input_shape=mdp.info.observation_space.shape,
output_shape=mdp.info.action_space.shape)
n_weights = approximator.weights_size
mu = np.zeros(n_weights)
sigma = 2e-0 * np.ones(n_weights)
policy = DeterministicPolicy(approximator)
dist = GaussianDiagonalDistribution(mu, sigma)
agent = alg(dist, policy, mdp.info, **params)
# Train
print(alg.__name__)
dataset_callback = CollectDataset()
core = Core(agent, mdp, callbacks=[dataset_callback])
for i in range(n_epochs):
core.learn(n_episodes=n_episodes,
n_episodes_per_fit=n_ep_per_fit,
render=False)
J = compute_J(dataset_callback.get(), gamma=mdp.info.gamma)
dataset_callback.clean()
p = dist.get_parameters()
print('mu: ', p[:n_weights])
print('sigma: ', p[n_weights:])
print('Reward at iteration ' + str(i) + ': ' + str(np.mean(J)))
print('Press a button to visualize the segway...')
input()
core.evaluate(n_episodes=3, render=True)
def experiment(policy, name, alg_version):
np.random.seed()
# MDP
if name == "Taxi":
mdp = generate_taxi('../grid.txt')
max_steps = 100000
evaluation_frequency = 2000
test_samples = 10000
elif name == "NChain-v0":
mdp = generate_chain(horizon=1000)
max_steps = 5000
evaluation_frequency = 100
test_samples = 10000
elif name == "Loop":
mdp = generate_loop(horizon=1000)
max_steps = 5000
evaluation_frequency = 100
test_samples = 10000
elif name == "SixArms":
mdp = generate_arms(horizon=1000)
max_steps = 25000
evaluation_frequency = 500
test_samples = 10000
elif name == "RiverSwim":
mdp = generate_river(horizon=1000)
max_steps = 5000
evaluation_frequency = 100
test_samples = 10000
else:
raise NotImplementedError
# Policy
# epsilon = ExponentialDecayParameter(value=1., decay_exp=.5,
# size=mdp.info.observation_space.size)
epsilon_train = ExponentialDecayParameter(
value=1., decay_exp=.5, size=mdp.info.observation_space.size)
epsilon_test = Parameter(0)
pi = policy(epsilon=epsilon_train)
# Agent
learning_rate = ExponentialDecayParameter(value=1.,
decay_exp=.2,
size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate)
agent = QLearning(pi, mdp.info, **algorithm_params)
# Algorithm
collect_dataset = CollectDataset()
callbacks = [collect_dataset]
core = Core(agent, mdp, callbacks)
scores = list()
scores_train = list()
# Train
for n_epoch in range(1, max_steps // evaluation_frequency + 1):
print('- Learning:')
# learning step
pi.set_epsilon(epsilon_train)
core.learn(n_steps=evaluation_frequency,
n_steps_per_fit=1,
quiet=False)
dataset = collect_dataset.get()
if name == "Taxi":
scores_train.append(get_stats(dataset))
elif name in ["SixArms"]:
scores_train.append(compute_scores_Loop(dataset, horizon=500))
else:
scores_train.append(compute_scores_Loop(dataset))
collect_dataset.clean()
mdp.reset()
print('- Evaluation:')
# evaluation step
pi.set_epsilon(epsilon_test)
dataset = core.evaluate(n_steps=test_samples, quiet=False)
mdp.reset()
scores.append(get_stats(dataset))
#np.save(env + '/'+alg_version+'_scores.npy', scores)
return scores_train, scores
def experiment(algorithm, name, update_mode, update_type, policy,
n_approximators, q_max, q_min, lr_exp, double, file_name,
out_dir, collect_qs, seed):
set_global_seeds(seed)
print('Using seed %s' % seed)
# MDP
if name == 'Taxi':
mdp = generate_taxi('../../grid.txt', horizon=5000)
max_steps = 500000
evaluation_frequency = 5000
test_samples = 5000
elif name == 'Chain':
mdp = generate_chain(horizon=100)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'Loop':
mdp = generate_loop(horizon=100)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'RiverSwim':
mdp = generate_river(horizon=100)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'SixArms':
mdp = generate_arms(horizon=100)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'KnightQuest':
mdp = Gym('KnightQuest-v0', gamma=0.99, horizon=10000)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
else:
raise NotImplementedError
epsilon_test = Parameter(0)
# Agent
learning_rate = ExponentialDecayParameter(value=1.,
decay_exp=lr_exp,
size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate)
if algorithm == 'ql':
if policy not in ['boltzmann', 'eps-greedy']:
warnings.warn(
'QL available with only boltzmann and eps-greedy policies!')
policy = 'eps-greedy'
if policy == 'eps-greedy':
epsilon_train = ExponentialDecayParameter(
value=1., decay_exp=.5, size=mdp.info.observation_space.size)
pi = policy_dict[policy](epsilon=epsilon_train)
else:
beta_train = ExponentialDecayParameter(
value=1.5 * q_max,
decay_exp=.5,
size=mdp.info.observation_space.size)
pi = policy_dict[policy](beta=beta_train)
if double:
agent = DoubleQLearning(pi, mdp.info, **algorithm_params)
else:
agent = QLearning(pi, mdp.info, **algorithm_params)
elif algorithm == 'boot-ql':
if policy not in ['boot', 'weighted']:
warnings.warn(
'Bootstrapped QL available with only boot and weighted policies!'
)
policy = 'boot'
pi = policy_dict[policy](n_approximators=n_approximators)
algorithm_params = dict(n_approximators=n_approximators,
mu=(q_max + q_min) / 2,
sigma=q_max - q_min,
**algorithm_params)
if double:
agent = BootstrappedDoubleQLearning(pi, mdp.info,
**algorithm_params)
else:
agent = BootstrappedQLearning(pi, mdp.info, **algorithm_params)
epsilon_train = Parameter(0)
elif algorithm == 'particle-ql':
if policy not in ['weighted', 'vpi']:
warnings.warn(
'Particle QL available with only vpi and weighted policies!')
policy = 'weighted'
pi = policy_dict[policy](n_approximators=n_approximators)
algorithm_params = dict(n_approximators=n_approximators,
update_mode=update_mode,
update_type=update_type,
q_max=q_max,
q_min=q_min,
**algorithm_params)
if double:
agent = ParticleDoubleQLearning(pi, mdp.info, **algorithm_params)
else:
agent = ParticleQLearning(pi, mdp.info, **algorithm_params)
epsilon_train = Parameter(0)
else:
raise ValueError()
# Algorithm
collect_dataset = CollectDataset()
collect_qs_callback = CollectQs(agent.approximator)
callbacks = [collect_dataset]
if collect_qs:
callbacks += [collect_qs_callback]
core = Core(agent, mdp, callbacks)
train_scores = []
test_scores = []
for n_epoch in range(1, max_steps // evaluation_frequency + 1):
# Train
if hasattr(pi, 'set_epsilon'):
pi.set_epsilon(epsilon_train)
if hasattr(pi, 'set_eval'):
pi.set_eval(False)
core.learn(n_steps=evaluation_frequency, n_steps_per_fit=1, quiet=True)
dataset = collect_dataset.get()
scores = compute_scores(dataset, mdp.info.gamma)
# print('Train: ', scores)
train_scores.append(scores)
collect_dataset.clean()
mdp.reset()
if hasattr(pi, 'set_epsilon'):
pi.set_epsilon(epsilon_test)
if hasattr(pi, 'set_eval'):
pi.set_eval(True)
dataset = core.evaluate(n_steps=test_samples, quiet=True)
mdp.reset()
scores = compute_scores(dataset, mdp.info.gamma)
# print('Evaluation: ', scores)
test_scores.append(scores)
if collect_qs:
qs = collect_qs_callback.get_values()
if not os.path.exists(out_dir):
os.makedirs(out_dir)
np.save(out_dir + '/' + file_name, qs)
return train_scores, test_scores
def experiment(n_epochs, n_episodes):
np.random.seed()
# MDP
n_steps = 5000
mdp = InvertedPendulum(horizon=n_steps)
# Agent
n_tilings = 11
alpha_r = Parameter(.0001)
alpha_theta = Parameter(.001 / n_tilings)
alpha_v = Parameter(.1 / n_tilings)
tilings = Tiles.generate(n_tilings-1, [10, 10],
mdp.info.observation_space.low,
mdp.info.observation_space.high + 1e-3)
phi = Features(tilings=tilings)
tilings_v = tilings + Tiles.generate(1, [1, 1],
mdp.info.observation_space.low,
mdp.info.observation_space.high + 1e-3)
psi = Features(tilings=tilings_v)
input_shape = (phi.size,)
mu = Regressor(LinearApproximator, input_shape=input_shape,
output_shape=mdp.info.action_space.shape)
std = Regressor(LinearApproximator, input_shape=input_shape,
output_shape=mdp.info.action_space.shape)
std_0 = np.sqrt(1.)
std.set_weights(np.log(std_0) / n_tilings * np.ones(std.weights_size))
policy = StateLogStdGaussianPolicy(mu, std)
agent = SAC_AVG(policy, mdp.info,
alpha_theta, alpha_v, alpha_r,
lambda_par=.5,
value_function_features=psi,
policy_features=phi)
# Train
dataset_callback = CollectDataset()
display_callback = Display(agent._V, mu, std,
mdp.info.observation_space.low,
mdp.info.observation_space.high,
phi, psi)
core = Core(agent, mdp, callbacks=[dataset_callback])
for i in range(n_epochs):
core.learn(n_episodes=n_episodes,
n_steps_per_fit=1, render=False)
J = compute_J(dataset_callback.get(), gamma=1.)
dataset_callback.clean()
display_callback()
print('Mean Reward at iteration ' + str(i) + ': ' +
str(np.sum(J) / n_steps/n_episodes))
print('Press a button to visualize the pendulum...')
input()
core.evaluate(n_steps=n_steps, render=True)
def experiment(algorithm,
name,
update_mode,
update_type,
policy,
n_approximators,
q_max,
q_min,
lr_exp,
file_name,
out_dir,
particles,
R=1,
m=1,
collect_qs=False,
seed=0):
set_global_seeds(seed)
print('Using seed %s' % seed)
# MDP
if name == 'Taxi':
mdp = generate_taxi('../../grid.txt', horizon=5000)
max_steps = 500000
evaluation_frequency = 5000
test_samples = 5000
elif name == 'Chain':
mdp = generate_chain(horizon=100)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'Loop':
mdp = generate_loop(horizon=100)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'RiverSwim':
mdp = generate_river(horizon=100)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'SixArms':
mdp = generate_arms(horizon=100)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'KnightQuest':
mdp = None
try:
mdp = Gym('KnightQuest-v0', gamma=0.99, horizon=10000)
except:
register(
id='KnightQuest-v0',
entry_point='envs.knight_quest:KnightQuest',
)
mdp = Gym('KnightQuest-v0', gamma=0.99, horizon=10000)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
else:
raise NotImplementedError
epsilon_test = Parameter(0)
# Agent
learning_rate = ExponentialDecayParameter(value=1.,
decay_exp=lr_exp,
size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate)
if algorithm == 'particle-ql':
delta = 0.1
if policy not in ['weighted', 'vpi', 'ucb']:
warnings.warn(
'Particle QL available with only vpi and weighted policies!')
policy = 'weighted'
if policy == 'ucb':
pi = UCBPolicy(delta=delta, q_max=R / (1 - mdp.info.gamma))
else:
pi = policy_dict[policy](n_approximators=n_approximators)
algorithm_params = dict(n_approximators=n_approximators,
update_mode=update_mode,
update_type=update_type,
q_max=q_max,
q_min=q_min,
delta=delta,
init_values=particles,
**algorithm_params)
agent = ParticleQLearning(pi, mdp.info, **algorithm_params)
if policy == 'ucb':
q = agent.approximator
quantiles = [
i * 1. / (n_approximators - 1) for i in range(n_approximators)
]
for p in range(n_approximators):
if quantiles[p] >= 1 - delta:
particle_bound = p
break
def quantile_func(state, quantile):
q_list = list()
for i in range(n_approximators):
q_list.append(q.predict(state, idx=i))
qs = np.array(q_list)
out = np.zeros(qs.shape[1])
out[:] = qs[particle_bound, :]
return out
def mu(state):
q_list = list()
for i in range(n_approximators):
q_list.append(q.predict(state, idx=i))
qs = np.array(q_list)
return np.mean(qs, axis=0)
pi.set_quantile_func(quantile_func)
pi.set_mu(mu)
epsilon_train = Parameter(0)
elif algorithm == 'delayed-ql':
algorithm_params = dict(R=R, m=m, **algorithm_params)
agent = DelayedQLearning(mdp.info, **algorithm_params)
pi = agent
# Algorithm
collect_dataset = CollectDataset()
callbacks = [collect_dataset]
if collect_qs:
collect_qs_callback = CollectQs(agent.approximator)
callbacks += [collect_qs_callback]
core = Core(agent, mdp, callbacks)
train_scores = []
test_scores = []
for n_epoch in range(1, max_steps // evaluation_frequency + 1):
# Train
if hasattr(pi, 'set_epsilon'):
pi.set_epsilon(epsilon_train)
if hasattr(pi, 'set_eval'):
pi.set_eval(False)
core.learn(n_steps=evaluation_frequency, n_steps_per_fit=1, quiet=True)
dataset = collect_dataset.get()
scores = compute_scores(dataset, mdp.info.gamma)
# print('Train: ', scores)
train_scores.append(scores)
collect_dataset.clean()
mdp.reset()
if hasattr(pi, 'set_epsilon'):
pi.set_epsilon(epsilon_test)
if hasattr(pi, 'set_eval'):
pi.set_eval(True)
dataset = core.evaluate(n_steps=test_samples, quiet=True)
mdp.reset()
scores = compute_scores(dataset, mdp.info.gamma)
# print('Evaluation: ', scores)
test_scores.append(scores)
if collect_qs:
qs = collect_qs_callback.get_values()
if not os.path.exists(out_dir):
os.makedirs(out_dir)
np.save(out_dir + '/' + file_name, qs)
return train_scores, test_scores
def experiment(algorithm, name, update_mode, update_type, policy, n_approximators, q_max, q_min,
lr_exp, R, log_lr, r_max_m, delayed_m, delayed_epsilon, delta, debug, double,
regret_test, a, b, mbie_C, value_iterations, tolerance, file_name, out_dir,
collect_qs, seed):
set_global_seeds(seed)
print('Using seed %s' % seed)
# MDP
if name == 'Taxi':
mdp = generate_taxi('../grid.txt', horizon=5000, gamma=0.99)
max_steps = 500000
evaluation_frequency = 5000
test_samples = 5000
elif name == 'Chain':
mdp = generate_chain(horizon=100, gamma=0.99)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'Gridworld':
mdp = generate_gridworld(horizon=100, gamma=0.99)
max_steps = 500000
evaluation_frequency = 5000
test_samples = 1000
elif name == 'Loop':
mdp = generate_loop(horizon=100, gamma=0.99)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'RiverSwim':
mdp = generate_river(horizon=100, gamma=0.99)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
mbie_C = 0.4
elif name == 'SixArms':
mdp = generate_arms(horizon=100, gamma=0.99)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
mbie_C = 0.8
elif name == 'ThreeArms':
horizon = 100
mdp = generate_three_arms(horizon=horizon, gamma=0.99)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
elif name == 'KnightQuest':
mdp = None
try:
mdp = Gym('KnightQuest-v0', gamma=0.99, horizon=10000)
except:
register(
id='KnightQuest-v0',
entry_point='envs.knight_quest:KnightQuest',
)
mdp = Gym('KnightQuest-v0', gamma=0.99, horizon=10000)
max_steps = 100000
evaluation_frequency = 1000
test_samples = 1000
else:
raise NotImplementedError
epsilon_test = Parameter(0)
# Agent
learning_rate = ExponentialDecayParameter(value=1., decay_exp=lr_exp,
size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate)
if regret_test:
max_steps = int(args.max_steps_regret * 1e6)
evaluation_frequency = max_steps // 100
test_samples = 1000
if debug:
max_steps = 100000
evaluation_frequency = max_steps // 100
test_samples = 1000
if algorithm == 'ql':
if policy not in ['boltzmann', 'eps-greedy']:
warnings.warn('QL available with only boltzmann and eps-greedy policies!')
policy = 'eps-greedy'
if policy == 'eps-greedy':
epsilon_train = ExponentialDecayParameter(value=1., decay_exp=lr_exp,
size=mdp.info.observation_space.size)
pi = policy_dict[policy](epsilon=epsilon_train)
else:
beta_train = ExponentialDecayParameter(value=1.5 * q_max, decay_exp=.5,
size=mdp.info.observation_space.size)
pi = policy_dict[policy](beta=beta_train)
if double:
agent = DoubleQLearning(pi, mdp.info, **algorithm_params)
else:
agent = QLearning(pi, mdp.info, **algorithm_params)
elif algorithm == 'boot-ql':
if policy not in ['boot', 'weighted']:
warnings.warn('Bootstrapped QL available with only boot and weighted policies!')
policy = 'boot'
pi = policy_dict[policy](n_approximators=n_approximators)
algorithm_params = dict(n_approximators=n_approximators,
mu=(q_max + q_min) / 2,
sigma=(q_max - q_min)/2,
**algorithm_params)
if double:
agent = BootstrappedDoubleQLearning(pi, mdp.info, **algorithm_params)
else:
agent = BootstrappedQLearning(pi, mdp.info, **algorithm_params)
epsilon_train = Parameter(0)
elif algorithm == 'particle-ql':
if policy not in ['weighted', 'ucb']:
warnings.warn('Particle QL available with only ucb and weighted policies!')
policy = 'weighted'
if policy == 'ucb':
pi = UCBPolicy(delta=delta, q_max=R/(1-mdp.info.gamma))
else:
pi = policy_dict[policy](n_approximators=n_approximators)
algorithm_params = dict(n_approximators=n_approximators,
update_mode=update_mode,
update_type=update_type,
q_max=q_max,
q_min=q_min,
delta=delta,
**algorithm_params)
if double:
agent = ParticleDoubleQLearning(pi, mdp.info, **algorithm_params)
else:
agent = ParticleQLearning(pi, mdp.info, **algorithm_params)
epsilon_train = Parameter(0)
elif algorithm == 'r-max':
thr_1 = int(np.ceil((4 * mdp.info.size[0] * 1.0/(1-mdp.info.gamma) * R )**3))
algorithm_params = dict(
rmax=R,
s_a_threshold=r_max_m
)
agent = RMaxAgent(mdp.info, **algorithm_params)
pi = agent
epsilon_train = Parameter(0)
elif algorithm == 'mbie':
algorithm_params = dict(
rmax=R,
C=mbie_C,
value_iterations=value_iterations,
tolerance=tolerance
)
agent = MBIE_EB(mdp.info, **algorithm_params)
pi = agent
epsilon_train = Parameter(0)
elif algorithm == 'delayed-ql':
theoretic_m = delayed_m
if regret_test:
gamma = mdp.info.gamma
Vmax = R / (1 - gamma)
epsilon = args.delayed_ratio * Vmax
delayed_epsilon = epsilon*(1-gamma)
delta = 0.1
S, A = mdp.info.size
theoretic_m = (1 + gamma*Vmax)**2 / (2*delayed_epsilon**2) * np.log(3*S*A/delta * (1 + S*A/(delayed_epsilon*(1-gamma))))
if debug:
print("Delta:{}".format(delta))
print("R:{}".format(R))
print("Vmax:{}".format(Vmax))
print("Gamma:{}".format(mdp.info.gamma))
print("Epsilon:{}".format(epsilon))
#print("k:{}".format(k))
print("m:{}".format(theoretic_m))
print("S:{}".format(S))
print("A:{}".format(A))
input()
def evaluate_policy(P, R, policy):
P_pi = np.zeros((S, S))
R_pi = np.zeros(S)
for s in range(S):
for s1 in range(S):
P_pi[s,s1] = np.sum(policy[s, :] * P[s, :, s1])
R_pi[s] = np.sum(policy[s, :] * np.sum(P[s, :, :] * R[s, :, :], axis=-1))
I = np.diag(np.ones(S))
V = np.linalg.solve(I - gamma * P_pi, R_pi)
return V
algorithm_params = dict(
R=R,
m=theoretic_m,
delta=delta,
epsilon=delayed_epsilon,
**algorithm_params)
agent = DelayedQLearning(mdp.info, **algorithm_params)
if regret_test:
collect_vs_callback = CollectVs(mdp, agent, evaluate_policy, args.freq_collection)
if debug:
print("Q:")
print(agent.get_approximator()[:, :])
print("Policy:")
print(agent.get_policy())
print("V:{}".format(evaluate_policy(mdp.p,mdp.r,agent.get_policy())))
input()
pi = agent
epsilon_train = Parameter(0)
elif algorithm == 'gaussian-ql':
if policy not in ['weighted-gaussian', 'ucb']:
warnings.warn('Particle QL available with only ucb and weighted policies!')
policy = 'weighted-gaussian'
if policy == 'ucb':
pi = UCBPolicy(delta=delta, q_max=R/(1-mdp.info.gamma))
else:
pi = policy_dict[policy]()
q_0 = (q_max - q_min) / 2
sigma_0 = (q_max - q_min) / np.sqrt(12)
C = 2 * R / (np.sqrt(2 * np.pi) * (1 - mdp.info.gamma) * sigma_0)
sigma_lr = None
if log_lr:
sigma_lr = LogarithmicDecayParameter(value=1., C=C,
size=mdp.info.size)
init_values = (q_0, sigma_0)
if regret_test:
sigma_lr = None
gamma = mdp.info.gamma
T = max_steps
S, A = mdp.info.size
a = (2 + gamma) / (2 *(1 - gamma))
b = a - 1
c = 1
d = b
q_max = R / (1 - gamma)
standard_bound = norm.ppf(1 - delta, loc=0, scale=1)
#first_fac = np.sqrt(b + T)
#second_fac = np.sqrt(a * np.log(S*A*T / delta))
#sigma2_factor = min(np.sqrt(b + T), np.sqrt(a * np.log(S*A*T / delta)))
q_0 = q_max
sigma1_0 = 0
#sigma2_0 = (R + gamma * q_max) / (standard_bound * np.sqrt(c-1)) * sigma2_factor
sigma2_0 = (gamma * q_max) / (c * standard_bound) * np.sqrt(a * np.log(S * A * T / delta))
init_values = (q_0, sigma1_0, sigma2_0)
learning_rate = TheoreticalParameter(a=a, b=b, decay_exp=1,
size=mdp.info.size)
learning_rate_sigma1 = TheoreticalParameter(a=a, b=b, decay_exp=1,
size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate,
sigma_1_learning_rate=learning_rate_sigma1)
sigma_lr = BetaParameter(c=c, d=d, size=mdp.info.size)
def evaluate_policy(P, R, policy):
P_pi = np.zeros((S, S))
R_pi = np.zeros(S)
for s in range(S):
for s1 in range(S):
P_pi[s,s1] = np.sum(policy[s, :] * P[s, :, s1])
R_pi[s] = np.sum(policy[s, :] * np.sum(P[s, :, :] * R[s, :, :],axis=-1))
I = np.diag(np.ones(S))
V = np.linalg.solve(I - gamma * P_pi, R_pi)
return V
if debug:
print("Delta:{}".format(delta))
print("R:{}".format(R))
print("Gamma:{}".format(mdp.info.gamma))
print("mu0:{}".format(q_0))
print("Sigma1_0:{}".format(sigma1_0))
print("Sigma2_0:{}".format(sigma2_0))
print("a:{}".format(a))
print("b:{}".format(b))
print("c:{}".format(c))
print("d:{}".format(d))
print("T:{}".format(T))
print("S:{}".format(S))
print("A:{}".format(A))
input()
algorithm_params = dict(
update_mode=update_mode,
update_type=update_type,
sigma_learning_rate=sigma_lr,
init_values=init_values,
delta=delta,
q_max=q_max,
**algorithm_params)
if double and not regret_test:
agent = GaussianDoubleQLearning(pi, mdp.info, **algorithm_params)
else:
agent = GaussianQLearning(pi, mdp.info, **algorithm_params)
if regret_test:
if debug:
freq = 10
else:
freq = args.freq_collection
collect_vs_callback = CollectVs(mdp, agent, evaluate_policy, freq)
if debug:
print("Policy:")
print(agent.get_policy())
print("Q")
for state in range(S):
means = np.array(agent.approximator.predict(np.array([state]), idx=0))
sigmas1 = np.array(agent.approximator.predict(np.array([state]), idx=1))
sigmas2 = np.array(agent.approximator.predict(np.array([state]), idx=2))
print("Means:{}".format(means))
print("Sigmas1:{}".format(sigmas1))
print("Sigmas2:{}".format(sigmas2))
print("V:{}".format(evaluate_policy(mdp.p,mdp.r,agent.get_policy())))
input()
if policy == 'ucb':
q = agent.approximator
standard_bound = norm.ppf(1 - delta, loc=0, scale=1)
def quantile_func(state):
means = np.array(q.predict(state, idx=0))
if regret_test:
sigmas1 = np.array(q.predict(state, idx=1))
sigmas2 = np.array(q.predict(state, idx=2))
sigmas = sigmas2
#print(sigmas1, sigmas2)
else:
sigmas = np.array(q.predict(state, idx=1))
out = sigmas * standard_bound + means
return out
def mu(state):
q_list = q.predict(state, idx=0)
means = np.array(q_list)
return means
pi.set_quantile_func(quantile_func)
pi.set_mu(mu)
epsilon_train = Parameter(0)
else:
raise ValueError()
# Algorithm
collect_dataset = CollectDataset()
callbacks = [collect_dataset]
if collect_qs:
if algorithm not in ['r-max']:
collect_qs_callback = CollectQs(agent.approximator)
callbacks += [collect_qs_callback]
if regret_test:
callbacks += [collect_vs_callback]
core = Core(agent, mdp, callbacks)
train_scores = []
test_scores = []
for n_epoch in range(1, max_steps // evaluation_frequency + 1):
# Train
if hasattr(pi, 'set_epsilon'):
pi.set_epsilon(epsilon_train)
if hasattr(pi, 'set_eval'):
pi.set_eval(False)
if regret_test:
collect_vs_callback.on()
core.learn(n_steps=evaluation_frequency, n_steps_per_fit=1, quiet=True)
dataset = collect_dataset.get()
scores = compute_scores(dataset, mdp.info.gamma)
#print('Train: ', scores)
train_scores.append(scores)
collect_dataset.clean()
mdp.reset()
if regret_test:
vs = collect_vs_callback.get_values()
if not os.path.exists(out_dir):
os.makedirs(out_dir)
print("Finished {} steps.".format(n_epoch * evaluation_frequency))
np.save(out_dir + "/vs_" + algorithm+"_"+str(seed), vs)
np.save(out_dir+"/scores_online" + str(seed), train_scores)
collect_vs_callback.off()
if hasattr(pi, 'set_epsilon'):
pi.set_epsilon(epsilon_test)
if hasattr(pi, 'set_eval'):
pi.set_eval(True)
dataset = core.evaluate(n_steps=test_samples, quiet=True)
s = mdp.reset()
scores = compute_scores(dataset, mdp.info.gamma)
print('Evaluation #%d:%s ' %(n_epoch, scores))
if debug:
print("Policy:")
print(agent.get_policy())
print("Q")
for state in range(S):
means = np.array(agent.approximator.predict(np.array([state]), idx=0))
sigmas1 = np.array(agent.approximator.predict(np.array([state]), idx=1))
sigmas2 = np.array(agent.approximator.predict(np.array([state]), idx=2))
print("Means:{}".format(means))
print("Sigmas1:{}".format(sigmas1))
print("Sigmas2:{}".format(sigmas2))
print("V:{}".format(evaluate_policy(mdp.p, mdp.r, agent.get_policy())))
input()
test_scores.append(scores)
if regret_test:
np.save(out_dir + "/scores_offline" + str(seed), test_scores)
if collect_qs:
qs= collect_qs_callback.get_values()
if not os.path.exists(out_dir):
os.makedirs(out_dir)
np.save(out_dir + '/' + file_name, qs)
return train_scores, test_scores