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 main(argv):
#env = retro.make(game='MegaMan-Nes', obs_type=retro.Observations.RAM)
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
retro.data.Integrations.add_custom_path(
os.path.join(SCRIPT_DIR, "integrations"))
print("megamanmain" in retro.data.list_games(
inttype=retro.data.Integrations.ALL))
env = RetroEnvironment('megamanmain', 5000, 0.9, obs_shape=256**3, \
obs_type=retro.Observations.RAM, \
use_restricted_actions=retro.Actions.DISCRETE,inttype=retro.data.Integrations.ALL)
epsilon = Parameter(value=1.)
learning_rate = Parameter(value=0.3)
policy = EpsGreedy(epsilon=epsilon)
agent = QLearning(env.info, policy, learning_rate)
#agent = CustomSARSA(env.info, policy, learning_rate)
core = Core(agent, env)
core.learn(n_episodes=50, n_steps_per_fit=1)
core.evaluate(n_episodes=2, render=True)
# print(agent.Q.shape, file=sys.stderr)
# shape = agent.Q.shape
# q = np.zeros(shape)
# for i in range(shape[0]):
# for j in range(shape[1]):
# state = np.array([i])
# action = np.array([j])
# q[i, j] = agent.Q.predict(state, action)
# print(q)
return 0
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 main(argv):
#env = retro.make(game='MegaMan-Nes', obs_type=retro.Observations.RAM)
env = RetroEnvironment('MegaMan-Nes', 5000, 0.9, obs_shape=256**3, \
obs_type=retro.Observations.RAM, \
use_restricted_actions=retro.Actions.DISCRETE)
epsilon = Parameter(value=1.)
learning_rate = Parameter(value=0.3)
policy = EpsGreedy(epsilon=epsilon)
agent = SARSA(env.info, policy, learning_rate)
#agent = CustomSARSA(env.info, policy, learning_rate)
core = Core(agent, env)
core.learn(n_steps=10000, n_steps_per_fit=1)
core.evaluate(n_episodes=10, render=True)
'''print(agent.Q.shape, file=sys.stderr)
shape = agent.Q.shape
q = np.zeros(shape)
for i in range(shape[0]):
for j in range(shape[1]):
state = np.array([i])
action = np.array([j])
q[i, j] = agent.Q.predict(state, action)
print(q)'''
return 0
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 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 test_dataset_utils():
np.random.seed(88)
mdp = GridWorld(3, 3, (2, 2))
epsilon = Parameter(value=0.)
alpha = Parameter(value=0.)
pi = EpsGreedy(epsilon=epsilon)
agent = SARSA(mdp.info, pi, alpha)
core = Core(agent, mdp)
dataset = core.evaluate(n_episodes=10)
J = compute_J(dataset, mdp.info.gamma)
J_test = np.array([
1.16106307e-03, 2.78128389e-01, 1.66771817e+00, 3.09031544e-01,
1.19725152e-01, 9.84770902e-01, 1.06111661e-02, 2.05891132e+00,
2.28767925e+00, 4.23911583e-01
])
assert np.allclose(J, J_test)
L = episodes_length(dataset)
L_test = np.array([87, 35, 18, 34, 43, 23, 66, 16, 15, 31])
assert np.array_equal(L, L_test)
dataset_ep = select_first_episodes(dataset, 3)
J = compute_J(dataset_ep, mdp.info.gamma)
assert np.allclose(J, J_test[:3])
L = episodes_length(dataset_ep)
assert np.allclose(L, L_test[:3])
samples = select_random_samples(dataset, 2)
s, a, r, ss, ab, last = parse_dataset(samples)
s_test = np.array([[6.], [1.]])
a_test = np.array([[0.], [1.]])
r_test = np.zeros(2)
ss_test = np.array([[3], [4]])
ab_test = np.zeros(2)
last_test = np.zeros(2)
assert np.array_equal(s, s_test)
assert np.array_equal(a, a_test)
assert np.array_equal(r, r_test)
assert np.array_equal(ss, ss_test)
assert np.array_equal(ab, ab_test)
assert np.array_equal(last, last_test)
index = np.sum(L_test[:2]) + L_test[2] // 2
min_J, max_J, mean_J, n_episodes = compute_metrics(dataset[:index],
mdp.info.gamma)
assert min_J == 0.0011610630703530948
assert max_J == 0.2781283894436937
assert mean_J == 0.1396447262570234
assert n_episodes == 2
def main(argv):
mdp = RetroFullEnvironment(
'MegaMan-Nes',
5000,
0.9,
#obs_type=retro.Observations.RAM,
use_restricted_actions=retro.Actions.DISCRETE)
epsilon = Parameter(value=1.)
learning_rate = Parameter(value=0.3)
train_frequency = 4
evaluation_frequency = 250000
target_update_frequency = 10000
initial_replay_size = 50000
#initial_replay_size = 500
max_replay_size = 500000
test_samples = 125000
max_steps = 50000000
policy = EpsGreedy(epsilon=epsilon)
optimizer = {'class': Adam, 'params': dict(lr=0.00025)}
approximator = KerasApproximator
approximator_params = dict(network=model,
input_shape=mdp.info.observation_space.shape,
output_shape=(mdp.info.action_space.n, ),
n_actions=mdp.info.action_space.n,
n_features=2048,
optimizer=optimizer,
loss=mean_squared_error,
print_summary=True)
algorithm_params = dict(batch_size=32,
target_update_frequency=target_update_frequency //
train_frequency,
replay_memory=None,
initial_replay_size=initial_replay_size,
max_replay_size=max_replay_size)
agent = DQN(mdp.info,
policy,
approximator,
approximator_params=approximator_params,
**algorithm_params)
core = Core(agent, mdp)
#core.learn(n_steps=1000000, n_steps_per_fit=1)
core.learn(n_steps=100000, n_steps_per_fit=1)
core.evaluate(n_episodes=10, render=True)
return 0
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 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 test_r_learning():
pi, mdp, _ = initialize()
agent = RLearning(mdp.info, pi, Parameter(.1), Parameter(.5))
core = Core(agent, mdp)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
test_q = np.array([[-6.19137991, -3.9368055, -5.11544257, -3.43673781],
[-2.52319391, 1.92201829, -2.77602918, -2.45972955],
[-5.38824415, -2.43019918, -1.09965936, 2.04202511],
[0., 0., 0., 0.]])
assert np.allclose(agent.Q.table, test_q)
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_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 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_boltzmann():
np.random.seed(88)
beta = Parameter(0.1)
pi = Boltzmann(beta)
Q = Table((10, 3))
Q.table = np.random.randn(10, 3)
pi.set_q(Q)
s = np.array([2])
a = np.array([1])
p_s = pi(s)
p_s_test = np.array([0.30676679, 0.36223227, 0.33100094])
assert np.allclose(p_s, p_s_test)
p_sa = pi(s, a)
p_sa_test = np.array([0.36223227])
assert np.allclose(p_sa, p_sa_test)
a = pi.draw_action(s)
a_test = 2
assert a.item() == a_test
beta_2 = LinearParameter(0.2, 0.1, 2)
pi.set_beta(beta_2)
p_sa_2 = pi(s, a)
assert p_sa_2 < p_sa
pi.update(s, a)
p_sa_3 = pi(s, a)
p_sa_3_test = np.array([0.33100094])
assert np.allclose(p_sa_3, p_sa_3_test)
def test_eps_greedy():
np.random.seed(88)
eps = Parameter(0.1)
pi = EpsGreedy(eps)
Q = Table((10, 3))
Q.table = np.random.randn(10, 3)
pi.set_q(Q)
s = np.array([2])
a = np.array([1])
p_s = pi(s)
p_s_test = np.array([0.03333333, 0.93333333, 0.03333333])
assert np.allclose(p_s, p_s_test)
p_sa = pi(s, a)
p_sa_test = np.array([0.93333333])
assert np.allclose(p_sa, p_sa_test)
a = pi.draw_action(s)
a_test = 1
assert a.item() == a_test
eps_2 = LinearParameter(0.2, 0.1, 2)
pi.set_epsilon(eps_2)
p_sa_2 = pi(s, a)
assert p_sa_2 < p_sa
pi.update(s, a)
pi.update(s, a)
p_sa_3 = pi(s, a)
print(eps_2.get_value())
assert p_sa_3 == p_sa
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_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 default(cls,
lr=.0001,
network=DQNNetwork,
initial_replay_size=50000,
max_replay_size=1000000,
batch_size=32,
target_update_frequency=2500,
n_steps_per_fit=1,
use_cuda=False):
policy = EpsGreedy(epsilon=Parameter(value=1.))
approximator_params = dict(network=network,
optimizer={
'class': optim.Adam,
'params': {
'lr': lr
}
},
loss=F.smooth_l1_loss,
use_cuda=use_cuda)
alg_params = dict(initial_replay_size=initial_replay_size,
max_replay_size=max_replay_size,
batch_size=batch_size,
target_update_frequency=target_update_frequency)
return cls(policy, TorchApproximator, approximator_params, alg_params,
n_steps_per_fit)
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_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 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 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_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 experiment():
np.random.seed()
# MDP
mdp = CartPole()
# Policy
epsilon = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
basis = [PolynomialBasis()]
s1 = np.array([-np.pi, 0, np.pi]) * .25
s2 = np.array([-1, 0, 1])
for i in s1:
for j in s2:
basis.append(GaussianRBF(np.array([i, j]), np.array([1.])))
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)
core.evaluate(n_episodes=3, render=True)
# Train
core.learn(n_episodes=100, n_episodes_per_fit=100)
# Test
test_epsilon = Parameter(0.)
agent.policy.set_epsilon(test_epsilon)
dataset = core.evaluate(n_episodes=1, quiet=True)
core.evaluate(n_steps=100, render=True)
return np.mean(episodes_length(dataset))
def test_copdac_q():
n_steps = 50
mdp = InvertedPendulum(horizon=n_steps)
np.random.seed(1)
torch.manual_seed(1)
torch.cuda.manual_seed(1)
# Agent
n_tilings = 1
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, [2, 2], 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
core = Core(agent, mdp)
core.learn(n_episodes=2, n_episodes_per_fit=1)
w = agent.policy.get_weights()
w_test = np.array([0, -6.62180045e-7, 0, -4.23972882e-2])
assert np.allclose(w, w_test)
def experiment(alpha):
np.random.seed()
# MDP
mdp = Gym(name='Acrobot-v1', horizon=np.inf, gamma=1.)
# Policy
epsilon = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
n_tilings = 10
tilings = Tiles.generate(n_tilings, [10, 10, 10, 10, 10, 10],
mdp.info.observation_space.low,
mdp.info.observation_space.high)
features = Features(tilings=tilings)
learning_rate = Parameter(alpha / n_tilings)
approximator_params = dict(input_shape=(features.size, ),
output_shape=(mdp.info.action_space.n, ),
n_actions=mdp.info.action_space.n)
algorithm_params = {'learning_rate': learning_rate, 'lambda_coeff': .9}
agent = TrueOnlineSARSALambda(mdp.info,
pi,
approximator_params=approximator_params,
features=features,
**algorithm_params)
#shape = agent.approximator.Q
#print(agent.Q)
# Algorithm
core = Core(agent, mdp)
# Train
core.learn(n_episodes=10, n_steps_per_fit=1, render=True)
dataset = core.evaluate(n_episodes=1, render=False)
#print(dataset)
print(episodes_length(dataset))
return np.mean(compute_J(dataset, .96))
def test_rq_learning_save(tmpdir):
agent_path = tmpdir / 'agent_{}'.format(datetime.now().strftime("%H%M%S%f"))
pi, mdp, _ = initialize()
agent_save = RQLearning(mdp.info, pi, Parameter(.1), beta=Parameter(.5))
core = Core(agent_save, mdp)
# 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 build(self, mdp_info):
self.approximator_params[
'input_shape'] = mdp_info.observation_space.shape
self.approximator_params['output_shape'] = (mdp_info.action_space.n, )
self.approximator_params['n_actions'] = mdp_info.action_space.n
self.epsilon = LinearParameter(value=1, threshold_value=.05, n=1000000)
self.epsilon_test = Parameter(value=.01)
return DoubleDQN(mdp_info, self.policy, self.approximator,
self.approximator_params, **self.alg_params)
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
