def test_sarsa_lambda_continuous_nn_save(tmpdir):
agent_path = tmpdir / 'agent_{}'.format(datetime.now().strftime("%H%M%S%f"))
pi, _, mdp_continuous = initialize()
mdp_continuous.seed(1)
features = Features(
n_outputs=mdp_continuous.info.observation_space.shape[0]
)
approximator_params = dict(
input_shape=(features.size,),
output_shape=(mdp_continuous.info.action_space.n,),
network=Network,
n_actions=mdp_continuous.info.action_space.n
)
agent_save = SARSALambdaContinuous(mdp_continuous.info, pi, TorchApproximator,
Parameter(.1), .9, features=features,
approximator_params=approximator_params)
core = Core(agent_save, mdp_continuous)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
agent_save.save(agent_path)
agent_load = Agent.load(agent_path)
for att, method in vars(agent_save).items():
save_attr = getattr(agent_save, att)
load_attr = getattr(agent_load, att)
tu.assert_eq(save_attr, load_attr)
def learn(alg, alg_params):
mdp = LQR.generate(dimensions=1)
np.random.seed(1)
torch.manual_seed(1)
torch.cuda.manual_seed(1)
approximator_params = dict(input_dim=mdp.info.observation_space.shape)
approximator = Regressor(LinearApproximator,
input_shape=mdp.info.observation_space.shape,
output_shape=mdp.info.action_space.shape,
params=approximator_params)
sigma = Regressor(LinearApproximator,
input_shape=mdp.info.observation_space.shape,
output_shape=mdp.info.action_space.shape,
params=approximator_params)
sigma_weights = 2 * np.ones(sigma.weights_size)
sigma.set_weights(sigma_weights)
policy = StateStdGaussianPolicy(approximator, sigma)
agent = alg(mdp.info, policy, **alg_params)
core = Core(agent, mdp)
core.learn(n_episodes=10, n_episodes_per_fit=5)
return policy
def experiment(alg, params, n_epochs, fit_per_run, ep_per_run):
np.random.seed()
# MDP
mdp = LQR.generate(dimensions=1)
approximator = Regressor(LinearApproximator,
input_shape=mdp.info.observation_space.shape,
output_shape=mdp.info.action_space.shape)
policy = DeterministicPolicy(mu=approximator)
mu = np.zeros(policy.weights_size)
sigma = 1e-3 * np.eye(policy.weights_size)
distribution = GaussianCholeskyDistribution(mu, sigma)
# Agent
agent = alg(mdp.info, distribution, policy, **params)
# Train
core = Core(agent, mdp)
dataset_eval = core.evaluate(n_episodes=ep_per_run)
print('distribution parameters: ', distribution.get_parameters())
J = compute_J(dataset_eval, gamma=mdp.info.gamma)
print('J at start : ' + str(np.mean(J)))
for i in range(n_epochs):
core.learn(n_episodes=fit_per_run * ep_per_run,
n_episodes_per_fit=ep_per_run)
dataset_eval = core.evaluate(n_episodes=ep_per_run)
print('distribution parameters: ', distribution.get_parameters())
J = compute_J(dataset_eval, gamma=mdp.info.gamma)
print('J at iteration ' + str(i) + ': ' + str(np.mean(J)))
def experiment():
np.random.seed()
# MDP
mdp = generate_simple_chain(state_n=5, goal_states=[2], prob=.8, rew=1,
gamma=.9)
# Policy
epsilon = Parameter(value=.15)
pi = EpsGreedy(epsilon=epsilon)
# Agent
learning_rate = Parameter(value=.2)
algorithm_params = dict(learning_rate=learning_rate)
agent = QLearning(mdp.info, pi, **algorithm_params)
# Core
core = Core(agent, mdp)
# Initial policy Evaluation
dataset = core.evaluate(n_steps=1000)
J = np.mean(compute_J(dataset, mdp.info.gamma))
print('J start:', J)
# Train
core.learn(n_steps=10000, n_steps_per_fit=1)
# Final Policy Evaluation
dataset = core.evaluate(n_steps=1000)
J = np.mean(compute_J(dataset, mdp.info.gamma))
print('J final:', J)
def learn_lspi():
np.random.seed(1)
# MDP
mdp = CartPole()
# Policy
epsilon = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
basis = [PolynomialBasis()]
features = Features(basis_list=basis)
fit_params = dict()
approximator_params = dict(input_shape=(features.size, ),
output_shape=(mdp.info.action_space.n, ),
n_actions=mdp.info.action_space.n)
agent = LSPI(mdp.info,
pi,
approximator_params=approximator_params,
fit_params=fit_params,
features=features)
# Algorithm
core = Core(agent, mdp)
# Train
core.learn(n_episodes=10, n_episodes_per_fit=10)
return agent
def experiment(algorithm_class, exp):
np.random.seed()
# MDP
p = np.load('chain_structure/p.npy')
rew = np.load('chain_structure/rew.npy')
mdp = FiniteMDP(p, rew, gamma=.9)
# Policy
epsilon = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
learning_rate = ExponentialParameter(value=1., exp=exp, size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate)
agent = algorithm_class(mdp.info, pi, **algorithm_params)
# Algorithm
collect_Q = CollectQ(agent.approximator)
callbacks = [collect_Q]
core = Core(agent, mdp, callbacks)
# Train
core.learn(n_steps=20000, n_steps_per_fit=1, quiet=True)
Qs = collect_Q.get()
return Qs
def test_sarsa_lambda_continuous_linear():
pi, _, mdp_continuous = initialize()
mdp_continuous.seed(1)
n_tilings = 1
tilings = Tiles.generate(n_tilings, [2, 2],
mdp_continuous.info.observation_space.low,
mdp_continuous.info.observation_space.high)
features = Features(tilings=tilings)
approximator_params = dict(
input_shape=(features.size,),
output_shape=(mdp_continuous.info.action_space.n,),
n_actions=mdp_continuous.info.action_space.n
)
agent = SARSALambdaContinuous(mdp_continuous.info, pi, LinearApproximator,
Parameter(.1), .9, features=features,
approximator_params=approximator_params)
core = Core(agent, mdp_continuous)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_w = np.array([-16.38428419, 0., -14.31250136, 0., -15.68571525, 0.,
-10.15663821, 0., -15.0545445, 0., -8.3683605, 0.])
assert np.allclose(agent.Q.get_weights(), test_w)
def test_sarsa_lambda_continuous_linear():
pi, _, mdp_continuous = initialize()
mdp_continuous.seed(1)
n_tilings = 1
tilings = Tiles.generate(n_tilings, [2, 2],
mdp_continuous.info.observation_space.low,
mdp_continuous.info.observation_space.high)
features = Features(tilings=tilings)
approximator_params = dict(
input_shape=(features.size, ),
output_shape=(mdp_continuous.info.action_space.n, ),
n_actions=mdp_continuous.info.action_space.n)
agent = SARSALambdaContinuous(mdp_continuous.info,
pi,
LinearApproximator,
Parameter(.1),
.9,
features=features,
approximator_params=approximator_params)
core = Core(agent, mdp_continuous)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_w = np.array([
-16.62627886, 0., -13.03033079, 0., -15.93237930, 0., -9.72299176, 0.,
-13.78884631, 0., -9.92157645, 0.
])
assert np.allclose(agent.Q.get_weights(), test_w)
def test_true_online_sarsa_lambda():
pi, _, mdp_continuous = initialize()
mdp_continuous.seed(1)
n_tilings = 1
tilings = Tiles.generate(n_tilings, [2, 2],
mdp_continuous.info.observation_space.low,
mdp_continuous.info.observation_space.high)
features = Features(tilings=tilings)
approximator_params = dict(
input_shape=(features.size, ),
output_shape=(mdp_continuous.info.action_space.n, ),
n_actions=mdp_continuous.info.action_space.n)
agent = TrueOnlineSARSALambda(mdp_continuous.info,
pi,
Parameter(.1),
.9,
features=features,
approximator_params=approximator_params)
core = Core(agent, mdp_continuous)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_w = np.array([
-17.30427303, 0., -13.54157504, 0., -16.82373134, 0., -10.29613337, 0.,
-14.79470382, 0., -10.50654665, 0.
])
assert np.allclose(agent.Q.get_weights(), test_w)
def learn(alg, alg_params):
mdp = InvertedPendulum(horizon=50)
np.random.seed(1)
torch.manual_seed(1)
torch.cuda.manual_seed(1)
critic_params = dict(network=Network,
optimizer={
'class': optim.Adam,
'params': {
'lr': 3e-4
}
},
loss=F.mse_loss,
input_shape=mdp.info.observation_space.shape,
output_shape=(1, ))
policy_params = dict(std_0=1., use_cuda=False)
policy = GaussianTorchPolicy(Network, mdp.info.observation_space.shape,
mdp.info.action_space.shape, **policy_params)
alg_params['critic_params'] = critic_params
agent = alg(mdp.info, policy, **alg_params)
core = Core(agent, mdp)
core.learn(n_episodes=2, n_episodes_per_fit=1)
return agent
def test_true_online_sarsa_lambda():
pi, _, mdp_continuous = initialize()
mdp_continuous.seed(1)
n_tilings = 1
tilings = Tiles.generate(n_tilings, [2, 2],
mdp_continuous.info.observation_space.low,
mdp_continuous.info.observation_space.high)
features = Features(tilings=tilings)
approximator_params = dict(
input_shape=(features.size,),
output_shape=(mdp_continuous.info.action_space.n,),
n_actions=mdp_continuous.info.action_space.n
)
agent = TrueOnlineSARSALambda(mdp_continuous.info, pi,
Parameter(.1), .9, features=features,
approximator_params=approximator_params)
core = Core(agent, mdp_continuous)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_w = np.array([-17.27410736, 0., -15.04386343, 0., -16.6551805, 0.,
-11.31383707, 0., -16.11782002, 0., -9.6927357, 0.])
assert np.allclose(agent.Q.get_weights(), test_w)
def experiment(goal, use_muscles, n_epochs, n_steps, n_episodes_test):
np.random.seed(1)
# MDP
gamma = 0.99
horizon = 2000
mdp = create_mdp(gamma, horizon, goal, use_muscles=use_muscles)
# Agent
agent = create_SAC_agent(mdp)
# normalization callback
normalizer = MinMaxPreprocessor(mdp_info=mdp.info)
# plotting callback
plotter = PlotDataset(mdp.info)
# Algorithm(with normalization and plotting)
core = Core(agent, mdp, callback_step=plotter, preprocessors=[normalizer])
# training loop
for n in range(n_epochs):
core.learn(n_steps=n_steps, n_steps_per_fit=1)
dataset = core.evaluate(n_episodes=n_episodes_test, render=True)
print('Epoch: ', n, ' J: ', np.mean(compute_J(dataset, gamma)),
' Len_ep: ', int(np.round(np.mean(episodes_length(dataset)))))
print('Press a button to visualize humanoid')
input()
core.evaluate(n_episodes=10, render=True)
def learn(alg, alg_params):
mdp = CarOnHill()
np.random.seed(1)
# Policy
epsilon = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon)
# Approximator
approximator_params = dict(input_shape=mdp.info.observation_space.shape,
n_actions=mdp.info.action_space.n,
n_estimators=50,
min_samples_split=5,
min_samples_leaf=2)
approximator = ExtraTreesRegressor
# Agent
agent = alg(mdp.info,
pi,
approximator,
approximator_params=approximator_params,
**alg_params)
# Algorithm
core = Core(agent, mdp)
# Train
core.learn(n_episodes=5, n_episodes_per_fit=5)
test_epsilon = Parameter(0.75)
agent.policy.set_epsilon(test_epsilon)
dataset = core.evaluate(n_episodes=2)
return agent, np.mean(compute_J(dataset, mdp.info.gamma))
def test_true_online_sarsa_lambda_save(tmpdir):
agent_path = tmpdir / 'agent_{}'.format(datetime.now().strftime("%H%M%S%f"))
pi, _, mdp_continuous = initialize()
mdp_continuous.seed(1)
n_tilings = 1
tilings = Tiles.generate(n_tilings, [2, 2],
mdp_continuous.info.observation_space.low,
mdp_continuous.info.observation_space.high)
features = Features(tilings=tilings)
approximator_params = dict(
input_shape=(features.size,),
output_shape=(mdp_continuous.info.action_space.n,),
n_actions=mdp_continuous.info.action_space.n
)
agent_save = TrueOnlineSARSALambda(mdp_continuous.info, pi,
Parameter(.1), .9, features=features,
approximator_params=approximator_params)
core = Core(agent_save, mdp_continuous)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
agent_save.save(agent_path)
agent_load = Agent.load(agent_path)
for att, method in vars(agent_save).items():
save_attr = getattr(agent_save, att)
load_attr = getattr(agent_load, att)
tu.assert_eq(save_attr, load_attr)
def test_collect_Q():
np.random.seed(88)
mdp = GridWorld(3, 3, (2, 2))
eps = Parameter(0.1)
pi = EpsGreedy(eps)
alpha = Parameter(0.1)
agent = SARSA(mdp.info, pi, alpha)
callback_q = CollectQ(agent.Q)
callback_max_q = CollectMaxQ(agent.Q, np.array([2]))
core = Core(agent, mdp, callbacks=[callback_q, callback_max_q])
core.learn(n_steps=1000, n_steps_per_fit=1, quiet=True)
V_test = np.array([2.4477574, 0.02246188, 1.6210059, 6.01867052])
V = callback_q.get()[-1]
assert np.allclose(V[0, :], V_test)
V_max = np.array([np.max(x[2, :], axis=-1) for x in callback_q.get()])
max_q = np.array(callback_max_q.get())
assert np.allclose(V_max, max_q)
def test_sarsa_lambda_continuous_nn():
pi, _, mdp_continuous = initialize()
mdp_continuous.seed(1)
features = Features(
n_outputs=mdp_continuous.info.observation_space.shape[0]
)
approximator_params = dict(
input_shape=(features.size,),
output_shape=(mdp_continuous.info.action_space.n,),
network=Network,
n_actions=mdp_continuous.info.action_space.n
)
agent = SARSALambdaContinuous(mdp_continuous.info, pi, TorchApproximator,
Parameter(.1), .9, features=features,
approximator_params=approximator_params)
core = Core(agent, mdp_continuous)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_w = np.array([-0.18968964, 0.4296857, 0.52967095, 0.5674884,
-0.12784956, -0.10572472, -0.14546978, -0.67001086,
-0.93925357])
assert np.allclose(agent.Q.get_weights(), test_w)
def test_lspi():
np.random.seed(1)
# MDP
mdp = CartPole()
# Policy
epsilon = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
basis = [PolynomialBasis()]
features = Features(basis_list=basis)
fit_params = dict()
approximator_params = dict(input_shape=(features.size, ),
output_shape=(mdp.info.action_space.n, ),
n_actions=mdp.info.action_space.n)
agent = LSPI(mdp.info,
pi,
approximator_params=approximator_params,
fit_params=fit_params,
features=features)
# Algorithm
core = Core(agent, mdp)
# Train
core.learn(n_episodes=10, n_episodes_per_fit=10)
w = agent.approximator.get_weights()
w_test = np.array([-1.00749128, -1.13444655, -0.96620322])
assert np.allclose(w, w_test)
def experiment(algorithm_class, exp):
np.random.seed()
# MDP
mdp = GridWorldVanHasselt()
# Policy
epsilon = ExponentialParameter(value=1,
exp=.5,
size=mdp.info.observation_space.size)
pi = EpsGreedy(epsilon=epsilon)
# Agent
learning_rate = ExponentialParameter(value=1, exp=exp, size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate)
agent = algorithm_class(mdp.info, pi, **algorithm_params)
# Algorithm
start = mdp.convert_to_int(mdp._start, mdp._width)
collect_max_Q = CollectMaxQ(agent.Q, start)
collect_dataset = CollectDataset()
callbacks = [collect_dataset, collect_max_Q]
core = Core(agent, mdp, callbacks)
# Train
core.learn(n_steps=10000, n_steps_per_fit=1, quiet=True)
_, _, reward, _, _, _ = parse_dataset(collect_dataset.get())
max_Qs = collect_max_Q.get()
return reward, max_Qs
def experiment(n_epochs, n_episodes):
np.random.seed()
logger = Logger(COPDAC_Q.__name__, results_dir=None)
logger.strong_line()
logger.info('Experiment Algorithm: ' + COPDAC_Q.__name__)
# 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(mdp.info,
policy,
mu,
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_fit=[dataset_callback])
for i in trange(n_epochs, leave=False):
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()
logger.epoch_info(i + 1, R_mean=np.sum(J) / n_steps / n_episodes)
logger.info('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(alg, env_id, horizon, gamma, n_epochs, n_steps, n_steps_per_fit,
n_episodes_test, alg_params, policy_params):
print(alg.__name__)
mdp = Gym(env_id, horizon, gamma)
critic_params = dict(network=Network,
optimizer={
'class': optim.Adam,
'params': {
'lr': 3e-4
}
},
loss=F.mse_loss,
n_features=32,
batch_size=64,
input_shape=mdp.info.observation_space.shape,
output_shape=(1, ))
policy = GaussianTorchPolicy(Network, mdp.info.observation_space.shape,
mdp.info.action_space.shape, **policy_params)
alg_params['critic_params'] = critic_params
agent = alg(mdp.info, policy, **alg_params)
core = Core(agent, mdp)
dataset = core.evaluate(n_episodes=n_episodes_test, render=False)
J = np.mean(compute_J(dataset, mdp.info.gamma))
R = np.mean(compute_J(dataset))
E = agent.policy.entropy()
tqdm.write('END OF EPOCH 0')
tqdm.write('J: {}, R: {}, entropy: {}'.format(J, R, E))
tqdm.write(
'##################################################################################################'
)
for it in trange(n_epochs):
core.learn(n_steps=n_steps, n_steps_per_fit=n_steps_per_fit)
dataset = core.evaluate(n_episodes=n_episodes_test, render=False)
J = np.mean(compute_J(dataset, mdp.info.gamma))
R = np.mean(compute_J(dataset))
E = agent.policy.entropy()
tqdm.write('END OF EPOCH ' + str(it + 1))
tqdm.write('J: {}, R: {}, entropy: {}'.format(J, R, E))
tqdm.write(
'##################################################################################################'
)
print('Press a button to visualize')
input()
core.evaluate(n_episodes=5, render=True)
def experiment(n_epochs, n_iterations, ep_per_run, save_states_to_disk):
np.random.seed()
logger = Logger('plot_and_norm_example', results_dir=None)
logger.strong_line()
logger.info('Plotting and normalization example')
# MDP
mdp = LQR.generate(dimensions=2, max_pos=10., max_action=5., episodic=True)
approximator = Regressor(LinearApproximator,
input_shape=mdp.info.observation_space.shape,
output_shape=mdp.info.action_space.shape)
sigma = Regressor(LinearApproximator,
input_shape=mdp.info.observation_space.shape,
output_shape=mdp.info.action_space.shape)
sigma_weights = 2 * np.ones(sigma.weights_size)
sigma.set_weights(sigma_weights)
policy = StateStdGaussianPolicy(approximator, sigma)
# Agent
optimizer = AdaptiveOptimizer(eps=.01)
algorithm_params = dict(optimizer=optimizer)
agent = REINFORCE(mdp.info, policy, **algorithm_params)
# normalization callback
prepro = MinMaxPreprocessor(mdp_info=mdp.info)
# plotting callback
plotter = PlotDataset(mdp.info, obs_normalized=True)
# Train
core = Core(agent, mdp, callback_step=plotter, preprocessors=[prepro])
# training loop
for n in range(n_epochs):
core.learn(n_episodes=n_iterations * ep_per_run,
n_episodes_per_fit=ep_per_run)
dataset = core.evaluate(n_episodes=ep_per_run, render=False)
J = np.mean(compute_J(dataset, mdp.info.gamma))
logger.epoch_info(n + 1, J=J)
if save_states_to_disk:
# save normalization / plot states to disk path
logger.info('Saving plotting and normalization data')
os.makedirs("./logs/plot_and_norm", exist_ok=True)
prepro.save("./logs/plot_and_norm/preprocessor.msh")
plotter.save_state("./logs/plot_and_norm/plotting_state")
# load states from disk path
logger.info('Loading preprocessor and plotter')
prerpo = MinMaxPreprocessor.load(
"./logs/plot_and_norm/preprocessor.msh")
plotter.load_state("./logs/plot_and_norm/plotting_state")
def experiment(alg, params, n_epochs, n_episodes, n_ep_per_fit):
np.random.seed()
print('============ start experiment ============')
logger = Logger(alg.__name__, results_dir=None)
logger.strong_line()
logger.info('Experiment Algorithm: ' + alg.__name__)
# MDP
mdp = GraspEnv()
print('============ mdp ============')
# Policy
n_weights = 6
mu = np.array([-0.5, 0.0, 0.91, m.pi, 0, 0])
sigma = np.asarray([0.05, 0.05, 0.05, 0.1, 0.1,
0.1]) #np.asarray([0.15, 0.15, 0.15, 0.4, 0.4, 0.4])
policy = Own_policy()
dist = GaussianDiagonalDistribution(
mu, sigma) # TODO: is this distribution right? Yes.
agent = alg(mdp.info, dist, policy, **params)
# Train
dataset_callback = CollectDataset(
) # TODO: should we also collect the dataset? Just keep this.
core = Core(agent, mdp, callbacks_fit=[dataset_callback])
#core = Core(agent, mdp)
for i in range(n_epochs):
print('================ core learn ================')
core.learn(n_episodes=n_episodes, n_episodes_per_fit=n_ep_per_fit)
J = compute_J(dataset_callback.get(), gamma=mdp.info.gamma)
print('J:', J)
print('============================')
dataset_callback.clean() # Todo: learning curve? Done
p = dist.get_parameters()
print('p:', p)
mu_0.append(p[:n_weights][0])
mu_1.append(p[:n_weights][1])
mu_2.append(p[:n_weights][2])
mu_3.append(p[:n_weights][3])
mu_4.append(p[:n_weights][4])
mu_5.append(p[:n_weights][5])
current_avg_sigma = (p[n_weights:][0] + p[n_weights:][1] +
p[n_weights:][2] + p[n_weights:][3] +
p[n_weights:][4] + p[n_weights:][5]) / 6
avg_sigma.append(current_avg_sigma)
# record learning curve of cumulative rewards
logger.epoch_info(i + 1,
J=np.mean(J),
mu=p[:n_weights],
sigma=p[n_weights:])
list_J.append(np.mean(J))
def learn(alg, alg_params):
# MDP
mdp = CartPole()
np.random.seed(1)
torch.manual_seed(1)
torch.cuda.manual_seed(1)
# Policy
epsilon_random = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon_random)
# Approximator
input_shape = mdp.info.observation_space.shape
approximator_params = dict(
network=Network if alg is not CategoricalDQN else FeatureNetwork,
optimizer={
'class': optim.Adam,
'params': {
'lr': .001
}
},
loss=F.smooth_l1_loss,
input_shape=input_shape,
output_shape=mdp.info.action_space.size,
n_actions=mdp.info.action_space.n,
n_features=2,
use_cuda=False)
# Agent
if alg not in [DuelingDQN, CategoricalDQN]:
agent = alg(mdp.info,
pi,
TorchApproximator,
approximator_params=approximator_params,
**alg_params)
elif alg is CategoricalDQN:
agent = alg(mdp.info,
pi,
approximator_params=approximator_params,
n_atoms=2,
v_min=-1,
v_max=1,
**alg_params)
else:
agent = alg(mdp.info,
pi,
approximator_params=approximator_params,
**alg_params)
# Algorithm
core = Core(agent, mdp)
core.learn(n_steps=500, n_steps_per_fit=5)
return agent
def learn(alg):
mdp = Gym('Pendulum-v0', 200, .99)
mdp.seed(1)
np.random.seed(1)
torch.manual_seed(1)
torch.cuda.manual_seed(1)
# Policy
policy_class = OrnsteinUhlenbeckPolicy
policy_params = dict(sigma=np.ones(1) * .2, theta=.15, dt=1e-2)
# Settings
initial_replay_size = 500
max_replay_size = 5000
batch_size = 200
n_features = 80
tau = .001
# Approximator
actor_input_shape = mdp.info.observation_space.shape
actor_params = dict(network=ActorNetwork,
n_features=n_features,
input_shape=actor_input_shape,
output_shape=mdp.info.action_space.shape,
use_cuda=False)
actor_optimizer = {'class': optim.Adam, 'params': {'lr': .001}}
critic_input_shape = (actor_input_shape[0] +
mdp.info.action_space.shape[0], )
critic_params = dict(network=CriticNetwork,
optimizer={
'class': optim.Adam,
'params': {
'lr': .001
}
},
loss=F.mse_loss,
n_features=n_features,
input_shape=critic_input_shape,
output_shape=(1, ),
use_cuda=False)
# Agent
agent = alg(mdp.info, policy_class, policy_params, actor_params,
actor_optimizer, critic_params, batch_size,
initial_replay_size, max_replay_size, tau)
# Algorithm
core = Core(agent, mdp)
core.learn(n_episodes=10, n_episodes_per_fit=5)
return agent.policy
def experiment():
np.random.seed()
# MDP
mdp = CarOnHill()
# Policy
epsilon = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon)
# Approximator
approximator_params = dict(input_shape=mdp.info.observation_space.shape,
n_actions=mdp.info.action_space.n,
n_estimators=50,
min_samples_split=5,
min_samples_leaf=2)
approximator = ExtraTreesRegressor
# Agent
algorithm_params = dict(n_iterations=20)
agent = FQI(mdp.info,
pi,
approximator,
approximator_params=approximator_params,
**algorithm_params)
# Algorithm
core = Core(agent, mdp)
# Render
core.evaluate(n_episodes=1, render=True)
# Train
core.learn(n_episodes=1000, n_episodes_per_fit=1000)
# Test
test_epsilon = Parameter(0.)
agent.policy.set_epsilon(test_epsilon)
initial_states = np.zeros((289, 2))
cont = 0
for i in range(-8, 9):
for j in range(-8, 9):
initial_states[cont, :] = [0.125 * i, 0.375 * j]
cont += 1
dataset = core.evaluate(initial_states=initial_states)
# Render
core.evaluate(n_episodes=3, render=True)
return np.mean(compute_J(dataset, mdp.info.gamma))
def experiment(alg, env_id, horizon, gamma, n_epochs, n_steps, n_steps_per_fit,
n_step_test, alg_params, policy_params):
logger = Logger(A2C.__name__, results_dir=None)
logger.strong_line()
logger.info('Experiment Algorithm: ' + A2C.__name__)
mdp = Gym(env_id, horizon, gamma)
critic_params = dict(network=Network,
optimizer={
'class': optim.RMSprop,
'params': {
'lr': 7e-4,
'eps': 1e-5
}
},
loss=F.mse_loss,
n_features=64,
batch_size=64,
input_shape=mdp.info.observation_space.shape,
output_shape=(1, ))
alg_params['critic_params'] = critic_params
policy = GaussianTorchPolicy(Network, mdp.info.observation_space.shape,
mdp.info.action_space.shape, **policy_params)
agent = alg(mdp.info, policy, **alg_params)
core = Core(agent, mdp)
dataset = core.evaluate(n_steps=n_step_test, render=False)
J = np.mean(compute_J(dataset, mdp.info.gamma))
R = np.mean(compute_J(dataset))
E = agent.policy.entropy()
logger.epoch_info(0, J=J, R=R, entropy=E)
for it in trange(n_epochs):
core.learn(n_steps=n_steps, n_steps_per_fit=n_steps_per_fit)
dataset = core.evaluate(n_steps=n_step_test, render=False)
J = np.mean(compute_J(dataset, mdp.info.gamma))
R = np.mean(compute_J(dataset))
E = agent.policy.entropy()
logger.epoch_info(it + 1, J=J, R=R, entropy=E)
logger.info('Press a button to visualize')
input()
core.evaluate(n_episodes=5, render=True)
def test_maxmin_q_learning():
pi, mdp, _ = initialize()
agent = MaxminQLearning(mdp.info, pi, Parameter(.5), n_tables=4)
core = Core(agent, mdp)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_q = np.array([[0., 0., 0., 0.], [0., 7.5, 0., 0.], [0., 0., 0., 5.],
[0., 0., 0., 0.]])
assert np.allclose(agent.Q[0].table, test_q)
def experiment(alg, params, n_epochs, fit_per_epoch, ep_per_fit):
np.random.seed()
logger = Logger(alg.__name__, results_dir=None)
logger.strong_line()
logger.info('Experiment Algorithm: ' + alg.__name__)
# MDP
mdp = ShipSteering()
# Policy
high = [150, 150, np.pi]
low = [0, 0, -np.pi]
n_tiles = [5, 5, 6]
low = np.array(low, dtype=np.float)
high = np.array(high, dtype=np.float)
n_tilings = 1
tilings = Tiles.generate(n_tilings=n_tilings,
n_tiles=n_tiles,
low=low,
high=high)
phi = Features(tilings=tilings)
input_shape = (phi.size, )
approximator = Regressor(LinearApproximator,
input_shape=input_shape,
output_shape=mdp.info.action_space.shape)
policy = DeterministicPolicy(approximator)
mu = np.zeros(policy.weights_size)
sigma = 4e-1 * np.ones(policy.weights_size)
distribution = GaussianDiagonalDistribution(mu, sigma)
# Agent
agent = alg(mdp.info, distribution, policy, features=phi, **params)
# Train
core = Core(agent, mdp)
dataset_eval = core.evaluate(n_episodes=ep_per_fit)
J = compute_J(dataset_eval, gamma=mdp.info.gamma)
logger.epoch_info(0, J=np.mean(J))
for i in range(n_epochs):
core.learn(n_episodes=fit_per_epoch * ep_per_fit,
n_episodes_per_fit=ep_per_fit)
dataset_eval = core.evaluate(n_episodes=ep_per_fit)
J = compute_J(dataset_eval, gamma=mdp.info.gamma)
logger.epoch_info(i + 1, J=np.mean(J))
def test_rq_learning():
pi, mdp, _ = initialize()
agent = RQLearning(mdp.info, pi, Parameter(.1), beta=Parameter(.5))
core = Core(agent, mdp)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_q = np.array([[0.32411217, 2.9698436, 0.46474438, 1.10269504],
[2.99505139, 5.217031, 0.40933461, 0.37687883],
[0.41942675, 0.32363486, 0., 4.68559],
[0., 0., 0., 0.]])
assert np.allclose(agent.Q.table, test_q)
agent = RQLearning(mdp.info, pi, Parameter(.1), delta=Parameter(.5))
core = Core(agent, mdp)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_q = np.array([[1.04081115e-2, 5.14662188e-1, 1.73951634e-2, 1.24081875e-01],
[0., 2.71, 1.73137500e-4, 4.10062500e-6],
[0., 4.50000000e-2, 0., 4.68559],
[0., 0., 0., 0.]])
assert np.allclose(agent.Q.table, test_q)
agent = RQLearning(mdp.info, pi, Parameter(.1), off_policy=True,
beta=Parameter(.5))
core = Core(agent, mdp)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_q = np.array([[3.55204022, 4.54235939, 3.42601165, 2.95170908],
[2.73877031, 3.439, 2.42031528, 2.86634531],
[3.43274708, 3.8592342, 3.72637395, 5.217031],
[0., 0., 0., 0.]])
assert np.allclose(agent.Q.table, test_q)
agent = RQLearning(mdp.info, pi, Parameter(.1), off_policy=True,
delta=Parameter(.5))
core = Core(agent, mdp)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_q = np.array([[0.18947806, 1.57782254, 0.21911489, 1.05197011],
[0.82309759, 5.217031, 0.04167492, 0.61472604],
[0.23620541, 0.59828262, 1.25299991, 5.217031],
[0., 0., 0., 0.]])
assert np.allclose(agent.Q.table, test_q)
def experiment(n_epochs, n_iterations, ep_per_run, save_states_to_disk):
np.random.seed()
# MDP
mdp = LQR.generate(dimensions=2, max_pos=10., max_action=5., episodic=True)
approximator = Regressor(LinearApproximator,
input_shape=mdp.info.observation_space.shape,
output_shape=mdp.info.action_space.shape)
sigma = Regressor(LinearApproximator,
input_shape=mdp.info.observation_space.shape,
output_shape=mdp.info.action_space.shape)
sigma_weights = 2 * np.ones(sigma.weights_size)
sigma.set_weights(sigma_weights)
policy = StateStdGaussianPolicy(approximator, sigma)
# Agent
learning_rate = AdaptiveParameter(value=.01)
algorithm_params = dict(learning_rate=learning_rate)
agent = REINFORCE(mdp.info, policy, **algorithm_params)
# normalization callback
prepro = MinMaxPreprocessor(mdp_info=mdp.info)
# plotting callback
plotter = PlotDataset(mdp.info, obs_normalized=True)
# Train
core = Core(agent, mdp, callback_step=plotter, preprocessors=[prepro])
# training loop
for n in range(n_epochs):
core.learn(n_episodes=n_iterations * ep_per_run,
n_episodes_per_fit=ep_per_run)
dataset = core.evaluate(n_episodes=ep_per_run, render=False)
print('Epoch: ', n, ' J: ', np.mean(compute_J(dataset,
mdp.info.gamma)))
if save_states_to_disk:
# save normalization / plot states to disk path
os.makedirs("./temp/", exist_ok=True)
prepro.save_state("./temp/normalization_state")
plotter.save_state("./temp/plotting_state")
# load states from disk path
prepro.load_state("./temp/normalization_state")
plotter.load_state("./temp/plotting_state")