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(pi, mdp.info, **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 experiment(algorithm_class, decay_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 = ExponentialDecayParameter(value=1.,
decay_exp=decay_exp,
size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate)
agent = algorithm_class(pi, mdp.info, **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_values()
return Qs
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(pi, mdp.info, 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 experiment_others(alg, decay_exp):
np.random.seed()
# MDP
grid_map = "simple_gridmap.txt"
mdp = GridWorldGenerator(grid_map=grid_map)
# Policy
epsilon = ExponentialDecayParameter(value=1, decay_exp=.5,
size=mdp.info.observation_space.size)
pi = EpsGreedy(epsilon=epsilon)
# Agent
alpha = ExponentialDecayParameter(value=1, decay_exp=decay_exp, size=mdp.info.size)
algorithm_params = dict(learning_rate=alpha)
fit_params = dict()
agent_params = {'algorithm_params': algorithm_params,
'fit_params': fit_params}
agent = alg(pi, mdp.info, agent_params)
# Algorithm
collect_max_Q = CollectMaxQ(agent.Q, mdp.convert_to_int(mdp._start, mdp._width))
collect_dataset = CollectDataset()
callbacks = [collect_max_Q, collect_dataset]
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_values()
return reward, max_Qs
def experiment2():
np.random.seed(3)
print('mushroom :')
# 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)
fit_params = dict()
agent_params = {
'algorithm_params': algorithm_params,
'fit_params': fit_params
}
agent = QLearning(pi, mdp.info, agent_params)
# Algorithm
collect_dataset = CollectDataset()
callbacks = [collect_dataset]
core = Core(agent, mdp, callbacks)
# Train
core.learn(n_steps=100, n_steps_per_fit=1, quiet=True)
dataset = collect_dataset.get()
return agent.Q.table
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)
fit_params = dict()
agent_params = {
'algorithm_params': algorithm_params,
'fit_params': fit_params
}
agent = QLearning(pi, mdp.info, agent_params)
# Algorithm
core = Core(agent, mdp)
# Train
core.learn(n_steps=10000, n_steps_per_fit=1)
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(approximator, pi, mdp.info,
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 np.mean(compute_J(dataset, mdp.info.gamma))
def test_lspi():
mdp = CartPole()
np.random.seed(1)
# Policy
epsilon = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
basis = [PolynomialBasis()]
features = Features(basis_list=basis)
approximator_params = dict(input_shape=(features.size, ),
output_shape=(mdp.info.action_space.n, ),
n_actions=mdp.info.action_space.n)
agent = LSPI(pi,
mdp.info,
fit_params=dict(),
approximator_params=approximator_params,
features=features)
# Algorithm
core = Core(agent, mdp)
# Train
core.learn(n_episodes=100, n_episodes_per_fit=100)
w = agent.approximator.get_weights()
w_test = np.array([-2.23880597, -2.27427603, -2.25])
assert np.allclose(w, w_test)
def experiment(boosted):
np.random.seed(20)
# MDP
mdp = CarOnHill()
# Policy
epsilon = Parameter(value=1)
pi = EpsGreedy(epsilon=epsilon)
# Approximator
if not boosted:
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)
else:
approximator_params = dict(
input_shape=mdp.info.observation_space.shape,
n_actions=mdp.info.action_space.n,
n_models=3,
prediction='sum',
n_estimators=50,
min_samples_split=5,
min_samples_leaf=2)
approximator = ExtraTreesRegressor
# Agent
algorithm_params = dict(n_iterations=3, boosted=boosted, quiet=True)
fit_params = dict()
agent_params = {
'approximator_params': approximator_params,
'algorithm_params': algorithm_params,
'fit_params': fit_params
}
agent = FQI(approximator, pi, mdp.info, agent_params)
# Algorithm
core = Core(agent, mdp)
# Train
core.learn(n_episodes=50, n_episodes_per_fit=50, quiet=True)
# Test
test_epsilon = Parameter(0)
agent.policy.set_epsilon(test_epsilon)
initial_states = np.zeros((9, 2))
cont = 0
for i in range(-8, 9, 8):
for j in range(-8, 9, 8):
initial_states[cont, :] = [0.125 * i, 0.375 * j]
cont += 1
dataset = core.evaluate(initial_states=initial_states, quiet=True)
return np.mean(compute_J(dataset, mdp.info.gamma))
def experiment(algorithm_class, decay_exp):
np.random.seed()
# MDP
mdp = GridWorldVanHasselt()
# Policy
epsilon = ExponentialDecayParameter(value=1, decay_exp=.5,
size=mdp.info.observation_space.size)
pi = EpsGreedy(epsilon=epsilon)
# Agent
learning_rate = ExponentialDecayParameter(value=1, decay_exp=decay_exp,
size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate)
agent = algorithm_class(pi, mdp.info, **algorithm_params)
# Algorithm
start = mdp.convert_to_int(mdp._start, mdp._width)
collect_max_Q = CollectMaxQ(agent.approximator, 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_values()
return reward, max_Qs
def build_high_level_agent(alg, params, mdp, epsilon):
pi = EpsGreedy(epsilon=epsilon, )
mdp_info_high = MDPInfo(observation_space=spaces.Discrete(16),
action_space=spaces.Discrete(4),
gamma=mdp.info.gamma,
horizon=100)
agent = alg(pi, mdp_info_high, **params)
return agent
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(pi, mdp.info, 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 == 0.0
assert max_J == 0.0011610630703530948
assert mean_J == 0.0005805315351765474
assert n_episodes == 2
def build_high_level_agent(alg, params, mdp):
epsilon = Parameter(value=0.1)
pi = EpsGreedy(epsilon=epsilon)
gamma = 1.0
mdp_info_agentH = MDPInfo(observation_space=spaces.Discrete(400),
action_space=spaces.Discrete(8),
gamma=gamma,
horizon=10000)
agent = alg(policy=pi, mdp_info=mdp_info_agentH, **params)
return agent
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 is not CategoricalDQN:
agent = alg(TorchApproximator,
pi,
mdp.info,
approximator_params=approximator_params,
**alg_params)
else:
agent = alg(pi,
mdp.info,
n_atoms=2,
v_min=-1,
v_max=1,
approximator_params=approximator_params,
**alg_params)
# Algorithm
core = Core(agent, mdp)
core.learn(n_steps=500, n_steps_per_fit=5)
return agent.approximator
def experiment1(decay_exp, beta_type):
np.random.seed()
# MDP
p = np.load('p.npy')
rew = np.load('rew.npy')
mdp = FiniteMDP(p, rew, gamma=.9)
# Policy
epsilon = Parameter(value=1)
pi = EpsGreedy(epsilon=epsilon)
# Agent
alpha = ExponentialDecayParameter(value=1,
decay_exp=decay_exp,
size=mdp.info.size)
if beta_type == 'Win':
beta = WindowedVarianceIncreasingParameter(value=1,
size=mdp.info.size,
tol=10.,
window=50)
else:
beta = VarianceIncreasingParameter(value=1,
size=mdp.info.size,
tol=10.)
algorithm_params = dict(learning_rate=alpha, beta=beta, off_policy=True)
fit_params = dict()
agent_params = {
'algorithm_params': algorithm_params,
'fit_params': fit_params
}
agent = RQLearning(pi, mdp.info, agent_params)
# Algorithm
collect_q = CollectQ(agent.Q)
collect_lr_1 = CollectParameters(beta, np.array([0]))
collect_lr_5 = CollectParameters(beta, np.array([4]))
callbacks = [collect_q, collect_lr_1, collect_lr_5]
core = Core(agent, mdp, callbacks)
# Train
core.learn(n_steps=20000, n_steps_per_fit=1, quiet=True)
Qs = collect_q.get_values()
lr_1 = collect_lr_1.get_values()
lr_5 = collect_lr_5.get_values()
return Qs, lr_1, lr_5
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(approximator, pi, mdp.info,
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(decay_exp, windowed, tol):
np.random.seed()
# MDP
mdp = GridWorldVanHasselt()
# Policy
epsilon = ExponentialDecayParameter(value=1,
decay_exp=.5,
size=mdp.info.observation_space.size)
pi = EpsGreedy(epsilon=epsilon)
# Agent
alpha = ExponentialDecayParameter(value=1,
decay_exp=decay_exp,
size=mdp.info.size)
if windowed:
beta = WindowedVarianceIncreasingParameter(value=1,
size=mdp.info.size,
tol=tol,
window=50)
else:
beta = VarianceIncreasingParameter(value=1,
size=mdp.info.size,
tol=tol)
algorithm_params = dict(learning_rate=alpha, beta=beta, off_policy=True)
fit_params = dict()
agent_params = {
'algorithm_params': algorithm_params,
'fit_params': fit_params
}
agent = RQLearning(pi, mdp.info, agent_params)
# Algorithm
collect_max_Q = CollectMaxQ(agent.Q,
mdp.convert_to_int(mdp._start, mdp._width))
collect_dataset = CollectDataset()
callbacks = [collect_max_Q, collect_dataset]
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_values()
return reward, max_Qs
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(pi,
mdp.info,
fit_params=fit_params,
approximator_params=approximator_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 experiment():
np.random.seed(3)
print('hierarchical :')
# MDP
mdp = generate_simple_chain(state_n=5,
goal_states=[2],
prob=.8,
rew=1,
gamma=.9)
# Model Block
model_block = MBlock(env=mdp, render=False)
# Policy
epsilon = Parameter(value=.15)
pi = EpsGreedy(epsilon=epsilon)
# Agent
learning_rate = Parameter(value=.2)
algorithm_params = dict(learning_rate=learning_rate)
fit_params = dict()
agent_params = {
'algorithm_params': algorithm_params,
'fit_params': fit_params
}
agent = QLearning(pi, mdp.info, agent_params)
# Control Block
control_block = ControlBlock(wake_time=1,
agent=agent,
n_eps_per_fit=None,
n_steps_per_fit=1)
# Algorithm
blocks = [model_block, control_block]
order = [0, 1]
model_block.add_input(control_block)
control_block.add_input(model_block)
control_block.add_reward(model_block)
computational_graph = ComputationalGraph(blocks=blocks, order=order)
core = HierarchicalCore(computational_graph)
# Train
core.learn(n_steps=100, quiet=True)
return agent.Q.table
def experiment():
np.random.seed(3)
print('hierarchical :')
mdp = GridWorldVanHasselt()
# Model Block
model_block = MBlock(env=mdp, render=False)
# Policy
epsilon = ExponentialDecayParameter(value=1,
decay_exp=.5,
size=mdp.info.observation_space.size)
pi = EpsGreedy(epsilon=epsilon)
# Agent
learning_rate = ExponentialDecayParameter(value=1.,
decay_exp=1.,
size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate)
fit_params = dict()
agent_params = {
'algorithm_params': algorithm_params,
'fit_params': fit_params
}
agent = QLearning(pi, mdp.info, agent_params)
# Control Block
control_block = ControlBlock(name='controller',
agent=agent,
n_steps_per_fit=1)
# Algorithm
blocks = [model_block, control_block]
order = [0, 1]
model_block.add_input(control_block)
control_block.add_input(model_block)
control_block.add_reward(model_block)
computational_graph = ComputationalGraph(blocks=blocks, order=order)
core = HierarchicalCore(computational_graph)
# Train
core.learn(n_steps=2000, quiet=True)
return agent.Q.table
def experiment(alpha):
gym.logger.setLevel(0)
np.random.seed(386)
# MDP
mdp = Gym(name='MountainCar-v0', horizon=10000, gamma=1.)
mdp.seed(201)
# Policy
epsilon = Parameter(value=0.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
learning_rate = Parameter(alpha)
tilings = Tiles.generate(10, [10, 10],
mdp.info.observation_space.low,
mdp.info.observation_space.high)
features = Features(tilings=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': .9}
fit_params = dict()
agent_params = {'approximator_params': approximator_params,
'algorithm_params': algorithm_params,
'fit_params': fit_params}
agent = TrueOnlineSARSALambda(pi, mdp.info, agent_params, features)
# Algorithm
core = Core(agent, mdp)
# Train
core.learn(n_steps=2000, n_steps_per_fit=1, quiet=True)
# Test
test_epsilon = Parameter(0.)
agent.policy.set_epsilon(test_epsilon)
initial_states = np.array([[0., 0.], [.1, .1]])
dataset = core.evaluate(initial_states=initial_states, quiet=True)
return np.mean(compute_J(dataset, 1.))
def test_collect_parameter():
np.random.seed(88)
mdp = GridWorld(3, 3, (2, 2))
eps = ExponentialParameter(value=1, exp=.5,
size=mdp.info.observation_space.size)
pi = EpsGreedy(eps)
alpha = Parameter(0.1)
agent = SARSA(pi, mdp.info, alpha)
callback_eps = CollectParameters(eps, 1)
core = Core(agent, mdp, callbacks=[callback_eps])
core.learn(n_steps=10, n_steps_per_fit=1, quiet=True)
eps_test = np.array([1., 0.70710678, 0.70710678, 0.57735027, 0.57735027,
0.57735027, 0.57735027, 0.57735027, 0.57735027, 0.57735027])
eps = callback_eps.get()
assert np.allclose(eps, eps_test)
def experiment():
np.random.seed()
# MDP
mdp = InvertedPendulum()
# Policy
epsilon = Parameter(value=0.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
rbfs = GaussianRBF.generate(10, [10, 10], mdp.info.observation_space.low,
mdp.info.observation_space.high)
features = Features(basis_list=rbfs)
approximator_params = dict(input_shape=(features.size, ),
output_shape=(mdp.info.action_space.n, ),
n_actions=mdp.info.action_space.n)
algorithm_params = dict(n_iterations)
fit_params = dict()
agent_params = {
'approximator_params': approximator_params,
'algorithm_params': algorithm_params,
'fit_params': fit_params
}
agent = LSPI(pi, mdp.info, agent_params, features)
# Algorithm
core = Core(agent, mdp)
# Train
core.learn(n_episodes=1000, n_episodes_per_fit=20)
# Test
test_epsilon = Parameter(0.)
agent.policy.set_epsilon(test_epsilon)
dataset = core.evaluate(n_episodes=20)
return np.mean(compute_J(dataset, 1.))
def experiment(alpha):
np.random.seed()
# MDP
mdp = Gym(name='MountainCar-v0', horizon=np.inf, gamma=1.)
# Policy
epsilon = Parameter(value=0.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
learning_rate = Parameter(alpha)
tilings = Tiles.generate(10, [10, 10], mdp.info.observation_space.low,
mdp.info.observation_space.high)
features = Features(tilings=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': .9}
fit_params = dict()
agent_params = {
'approximator_params': approximator_params,
'algorithm_params': algorithm_params,
'fit_params': fit_params
}
agent = TrueOnlineSARSALambda(pi, mdp.info, agent_params, features)
# Algorithm
collect_dataset = CollectDataset()
callbacks = [collect_dataset]
core = Core(agent, mdp, callbacks=callbacks)
# Train
core.learn(n_episodes=20, n_steps_per_fit=1, render=0)
dataset = collect_dataset.get()
return np.mean(compute_J(dataset, 1.))
def experiment(alpha):
np.random.seed()
# MDP
mdp = Gym(name='MountainCar-v0', horizon=np.inf, gamma=1.)
# Policy
epsilon = Parameter(value=0.)
pi = EpsGreedy(epsilon=epsilon)
# Agent
n_tilings = 10
tilings = Tiles.generate(n_tilings, [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(pi,
mdp.info,
approximator_params=approximator_params,
features=features,
**algorithm_params)
# Algorithm
core = Core(agent, mdp)
# Train
core.learn(n_episodes=40, n_steps_per_fit=1, render=False)
dataset = core.evaluate(n_episodes=1, render=True)
return np.mean(compute_J(dataset, 1.))
def test_collect_dataset():
np.random.seed(88)
callback = CollectDataset()
mdp = GridWorld(4, 4, (2, 2))
eps = Parameter(0.1)
pi = EpsGreedy(eps)
alpha = Parameter(0.2)
agent = SARSA(pi, mdp.info, alpha)
core = Core(agent, mdp, callbacks=[callback])
core.learn(n_steps=10, n_steps_per_fit=1, quiet=True)
dataset = callback.get()
assert len(dataset) == 10
core.learn(n_steps=5, n_steps_per_fit=1, quiet=True)
assert len(dataset) == 15
callback.clean()
dataset = callback.get()
assert len(dataset) == 0
def experiment2():
np.random.seed(3)
print('mushroom :')
mdp = GridWorldVanHasselt()
# Policy
epsilon = ExponentialDecayParameter(value=1,
decay_exp=.5,
size=mdp.info.observation_space.size)
pi = EpsGreedy(epsilon=epsilon)
# Agent
learning_rate = ExponentialDecayParameter(value=1.,
decay_exp=1.,
size=mdp.info.size)
algorithm_params = dict(learning_rate=learning_rate)
fit_params = dict()
agent_params = {
'algorithm_params': algorithm_params,
'fit_params': fit_params
}
agent = QLearning(pi, mdp.info, agent_params)
# Algorithm
collect_dataset = CollectDataset()
callbacks = [collect_dataset]
core = Core(agent, mdp, callbacks)
# Train
core.learn(n_steps=2000, n_steps_per_fit=1, quiet=True)
# Train
dataset = collect_dataset.get()
VisualizeControlBlock(dataset)
return agent.Q.table
def experiment():
np.random.seed()
# Argument parser
parser = argparse.ArgumentParser()
arg_game = parser.add_argument_group('Game')
arg_game.add_argument("--name",
type=str,
default='BreakoutDeterministic-v4',
help='Gym ID of the Atari game.')
arg_game.add_argument("--screen-width", type=int, default=84,
help='Width of the game screen.')
arg_game.add_argument("--screen-height", type=int, default=84,
help='Height of the game screen.')
arg_mem = parser.add_argument_group('Replay Memory')
arg_mem.add_argument("--initial-replay-size", type=int, default=50000,
help='Initial size of the replay memory.')
arg_mem.add_argument("--max-replay-size", type=int, default=500000,
help='Max size of the replay memory.')
arg_net = parser.add_argument_group('Deep Q-Network')
arg_net.add_argument("--optimizer",
choices=['adadelta',
'adam',
'rmsprop',
'rmspropcentered'],
default='adam',
help='Name of the optimizer to use to learn.')
arg_net.add_argument("--learning-rate", type=float, default=.00025,
help='Learning rate value of the optimizer. Only used'
'in rmspropcentered')
arg_net.add_argument("--decay", type=float, default=.95,
help='Discount factor for the history coming from the'
'gradient momentum in rmspropcentered')
arg_net.add_argument("--epsilon", type=float, default=.01,
help='Epsilon term used in rmspropcentered')
arg_alg = parser.add_argument_group('Algorithm')
arg_alg.add_argument("--algorithm", choices=['dqn', 'ddqn', 'adqn'],
default='dqn',
help='Name of the algorithm. dqn is for standard'
'DQN, ddqn is for Double DQN and adqn is for'
'Averaged DQN.')
arg_alg.add_argument("--n-approximators", type=int, default=1,
help="Number of approximators used in the ensemble for"
"Averaged DQN.")
arg_alg.add_argument("--batch-size", type=int, default=32,
help='Batch size for each fit of the network.')
arg_alg.add_argument("--history-length", type=int, default=4,
help='Number of frames composing a state.')
arg_alg.add_argument("--target-update-frequency", type=int, default=10000,
help='Number of learning step before each update of'
'the target network.')
arg_alg.add_argument("--evaluation-frequency", type=int, default=250000,
help='Number of learning step before each evaluation.'
'This number represents an epoch.')
arg_alg.add_argument("--train-frequency", type=int, default=4,
help='Number of learning steps before each fit of the'
'neural network.')
arg_alg.add_argument("--max-steps", type=int, default=50000000,
help='Total number of learning steps.')
arg_alg.add_argument("--final-exploration-frame", type=int, default=1000000,
help='Number of steps until the exploration rate stops'
'decreasing.')
arg_alg.add_argument("--initial-exploration-rate", type=float, default=1.,
help='Initial value of the exploration rate.')
arg_alg.add_argument("--final-exploration-rate", type=float, default=.1,
help='Final value of the exploration rate. When it'
'reaches this values, it stays constant.')
arg_alg.add_argument("--test-exploration-rate", type=float, default=.05,
help='Exploration rate used during evaluation.')
arg_alg.add_argument("--test-samples", type=int, default=125000,
help='Number of steps for each evaluation.')
arg_alg.add_argument("--max-no-op-actions", type=int, default=8,
help='Maximum number of no-op action performed at the'
'beginning of the episodes. The minimum number is'
'history_length. This number is 30 in the DQN'
'Deepmind paper, but they consider the first 30'
'frame without frame skipping.')
arg_alg.add_argument("--no-op-action-value", type=int, default=0,
help='Value of the no-op action.')
arg_utils = parser.add_argument_group('Utils')
arg_utils.add_argument('--load-path', type=str,
help='Path of the model to be loaded.')
arg_utils.add_argument('--save', action='store_true',
help='Flag specifying whether to save the model.')
arg_utils.add_argument('--render', action='store_true',
help='Flag specifying whether to render the game.')
arg_utils.add_argument('--quiet', action='store_true',
help='Flag specifying whether to hide the progress'
'bar.')
arg_utils.add_argument('--debug', action='store_true',
help='Flag specifying whether the script has to be'
'run in debug mode.')
args = parser.parse_args()
scores = list()
# Evaluation of the model provided by the user.
if args.load_path:
# MDP
mdp = Atari(args.name, args.screen_width, args.screen_height,
ends_at_life=False)
# Policy
epsilon_test = Parameter(value=args.test_exploration_rate)
pi = EpsGreedy(epsilon=epsilon_test)
# Approximator
input_shape = (args.screen_height, args.screen_width,
args.history_length)
approximator_params = dict(
input_shape=input_shape,
output_shape=(mdp.info.action_space.n,),
n_actions=mdp.info.action_space.n,
name='test',
load_path=args.load_path,
optimizer={'name': args.optimizer,
'lr': args.learning_rate,
'decay': args.decay,
'epsilon': args.epsilon}
)
approximator = ConvNet
# Agent
algorithm_params = dict(
batch_size=1,
train_frequency=1,
target_update_frequency=1,
initial_replay_size=0,
max_replay_size=0,
history_length=args.history_length,
max_no_op_actions=args.max_no_op_actions,
no_op_action_value=args.no_op_action_value,
dtype=np.uint8
)
agent = DQN(approximator, pi, mdp.info,
approximator_params=approximator_params, **algorithm_params)
# Algorithm
core_test = Core(agent, mdp)
# Evaluate model
pi.set_epsilon(epsilon_test)
dataset = core_test.evaluate(n_steps=args.test_samples,
render=args.render,
quiet=args.quiet)
get_stats(dataset)
else:
# DQN learning run
# Summary folder
folder_name = './logs/atari_' + args.algorithm + '_' + args.name +\
'_' + datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
# Settings
if args.debug:
initial_replay_size = 50
max_replay_size = 500
train_frequency = 5
target_update_frequency = 10
test_samples = 20
evaluation_frequency = 50
max_steps = 1000
else:
initial_replay_size = args.initial_replay_size
max_replay_size = args.max_replay_size
train_frequency = args.train_frequency
target_update_frequency = args.target_update_frequency
test_samples = args.test_samples
evaluation_frequency = args.evaluation_frequency
max_steps = args.max_steps
# MDP
mdp = Atari(args.name, args.screen_width, args.screen_height,
ends_at_life=True)
# Policy
epsilon = LinearDecayParameter(value=args.initial_exploration_rate,
min_value=args.final_exploration_rate,
n=args.final_exploration_frame)
epsilon_test = Parameter(value=args.test_exploration_rate)
epsilon_random = Parameter(value=1)
pi = EpsGreedy(epsilon=epsilon_random)
# Approximator
input_shape = (args.screen_height, args.screen_width,
args.history_length)
approximator_params = dict(
input_shape=input_shape,
output_shape=(mdp.info.action_space.n,),
n_actions=mdp.info.action_space.n,
folder_name=folder_name,
optimizer={'name': args.optimizer,
'lr': args.learning_rate,
'decay': args.decay,
'epsilon': args.epsilon}
)
approximator = ConvNet
# Agent
algorithm_params = dict(
batch_size=args.batch_size,
n_approximators=args.n_approximators,
initial_replay_size=initial_replay_size,
max_replay_size=max_replay_size,
history_length=args.history_length,
train_frequency=train_frequency,
target_update_frequency=target_update_frequency,
max_no_op_actions=args.max_no_op_actions,
no_op_action_value=args.no_op_action_value,
dtype=np.uint8
)
if args.algorithm == 'dqn':
agent = DQN(approximator, pi, mdp.info,
approximator_params=approximator_params,
**algorithm_params)
elif args.algorithm == 'ddqn':
agent = DoubleDQN(approximator, pi, mdp.info,
approximator_params=approximator_params,
**algorithm_params)
elif args.algorithm == 'adqn':
agent = AveragedDQN(approximator, pi, mdp.info,
approximator_params=approximator_params,
**algorithm_params)
# Algorithm
core = Core(agent, mdp)
# RUN
# Fill replay memory with random dataset
print_epoch(0)
core.learn(n_steps=initial_replay_size,
n_steps_per_fit=initial_replay_size, quiet=args.quiet)
if args.save:
agent.approximator.model.save()
# Evaluate initial policy
pi.set_epsilon(epsilon_test)
if args.algorithm == 'ddqn':
agent.policy.set_q(agent.target_approximator)
mdp.set_episode_end(False)
dataset = core.evaluate(n_steps=test_samples, render=args.render,
quiet=args.quiet)
scores.append(get_stats(dataset))
if args.algorithm == 'ddqn':
agent.policy.set_q(agent.approximator)
np.save(folder_name + '/scores.npy', scores)
for n_epoch in range(1, max_steps // evaluation_frequency + 1):
print_epoch(n_epoch)
print('- Learning:')
# learning step
pi.set_epsilon(epsilon)
mdp.set_episode_end(True)
core.learn(n_steps=evaluation_frequency,
n_steps_per_fit=train_frequency, quiet=args.quiet)
if args.save:
agent.approximator.model.save()
print('- Evaluation:')
# evaluation step
pi.set_epsilon(epsilon_test)
if args.algorithm == 'ddqn':
agent.policy.set_q(agent.target_approximator)
mdp.set_episode_end(False)
dataset = core.evaluate(n_steps=test_samples, render=args.render,
quiet=args.quiet)
scores.append(get_stats(dataset))
if args.algorithm == 'ddqn':
agent.policy.set_q(agent.approximator)
np.save(folder_name + '/scores.npy', scores)
return scores
def experiment_ghavamzade(alg_high, alg_low, params, subdir, i):
np.random.seed()
# Model Block
mdp = ShipSteering(small=False, n_steps_action=3)
#State Placeholder
state_ph = PlaceHolder(name='state_ph')
#Reward Placeholder
reward_ph = PlaceHolder(name='reward_ph')
#Last action Placeholder
lastaction_ph = PlaceHolder(name='lastaction_ph')
# FeaturesH
low_hi = 0
lim_hi = 1000 + 1e-8
n_tiles_high = [20, 20]
n_tilings = 1
# Discretization Block
discretization_block = DiscretizationBlock(low=low_hi,
high=lim_hi,
n_tiles=n_tiles_high)
# PolicyH
epsilon = Parameter(value=0.1)
piH = EpsGreedy(epsilon=epsilon)
# AgentH
learning_rate = params.get('learning_rate_high')
mdp_info_agentH = MDPInfo(observation_space=spaces.Discrete(
n_tiles_high[0] * n_tiles_high[1]),
action_space=spaces.Discrete(8),
gamma=1,
horizon=10000)
agentH = alg_high(policy=piH,
mdp_info=mdp_info_agentH,
learning_rate=learning_rate,
lambda_coeff=0.9)
epsilon_update = EpsilonUpdate(piH)
# Control Block H
control_blockH = ControlBlock(name='control block H',
agent=agentH,
n_steps_per_fit=1)
#FeaturesL
high = [150, 150, np.pi]
low = [0, 0, -np.pi]
n_tiles = [5, 5, 10]
low = np.array(low, dtype=np.float)
high = np.array(high, dtype=np.float)
n_tilings = 3
tilingsL = Tiles.generate(n_tilings=n_tilings,
n_tiles=n_tiles,
low=low,
high=high)
featuresL = Features(tilings=tilingsL)
mdp_info_agentL = MDPInfo(observation_space=spaces.Box(
low=np.array([0, 0]), high=np.array([150, 150]), shape=(2, )),
action_space=mdp.info.action_space,
gamma=0.99,
horizon=10000)
# Approximators
input_shape = (featuresL.size, )
approximator_params = dict(input_dim=input_shape[0])
approximator1 = Regressor(LinearApproximator,
input_shape=input_shape,
output_shape=mdp.info.action_space.shape,
**approximator_params)
approximator2 = Regressor(LinearApproximator,
input_shape=input_shape,
output_shape=mdp.info.action_space.shape,
**approximator_params)
# Policy1
std1 = np.array([3e-2])
pi1 = DiagonalGaussianPolicy(mu=approximator1, std=std1)
# Policy2
std2 = np.array([3e-2])
pi2 = DiagonalGaussianPolicy(mu=approximator2, std=std2)
# Agent1
learning_rate1 = params.get('learning_rate_low')
agent1 = alg_low(pi1, mdp_info_agentL, learning_rate1, featuresL)
# Agent2
learning_rate2 = params.get('learning_rate_low')
agent2 = alg_low(pi2, mdp_info_agentL, learning_rate2, featuresL)
#Termination Conds
termination_condition1 = TerminationCondition(active_dir='+')
termination_condition2 = TerminationCondition(active_dir='x')
low_ep_per_fit = params.get('low_ep_per_fit')
# Control Block +
control_block_plus = ControlBlock(
name='control block 1',
agent=agent1,
n_eps_per_fit=low_ep_per_fit,
termination_condition=termination_condition1)
# Control Block x
control_block_cross = ControlBlock(
name='control block 2',
agent=agent2,
n_eps_per_fit=low_ep_per_fit,
termination_condition=termination_condition2)
# Function Block 1: picks state for hi lev ctrl
function_block1 = fBlock(phi=pick_state, name='f1 pickstate')
# Function Block 2: maps the env to low lev ctrl state
function_block2 = fBlock(phi=rototranslate, name='f2 rotot')
# Function Block 3: holds curr state as ref
function_block3 = hold_state(name='f3 holdstate')
# Function Block 4: adds hi lev rew
function_block4 = addBlock(name='f4 add')
# Function Block 5: adds low lev rew
function_block5 = addBlock(name='f5 add')
# Function Block 6:ext rew of hi lev ctrl
function_block6 = fBlock(phi=G_high, name='f6 G_hi')
# Function Block 7: ext rew of low lev ctrl
function_block7 = fBlock(phi=G_low, name='f7 G_lo')
#Reward Accumulator H:
reward_acc_H = reward_accumulator_block(gamma=mdp_info_agentH.gamma,
name='reward_acc_H')
# Selector Block
function_block8 = fBlock(phi=selector_function, name='f7 G_lo')
#Mux_Block
mux_block = MuxBlock(name='mux')
mux_block.add_block_list([control_block_plus])
mux_block.add_block_list([control_block_cross])
#Algorithm
blocks = [
state_ph, reward_ph, lastaction_ph, control_blockH, mux_block,
function_block1, function_block2, function_block3, function_block4,
function_block5, function_block6, function_block7, function_block8,
reward_acc_H, discretization_block
]
reward_acc_H.add_input(reward_ph)
reward_acc_H.add_alarm_connection(control_block_plus)
reward_acc_H.add_alarm_connection(control_block_cross)
control_blockH.add_input(discretization_block)
control_blockH.add_reward(function_block4)
control_blockH.add_alarm_connection(control_block_plus)
control_blockH.add_alarm_connection(control_block_cross)
mux_block.add_input(function_block8)
mux_block.add_input(function_block2)
control_block_plus.add_reward(function_block5)
control_block_cross.add_reward(function_block5)
function_block1.add_input(state_ph)
function_block2.add_input(control_blockH)
function_block2.add_input(state_ph)
function_block2.add_input(function_block3)
function_block3.add_input(state_ph)
function_block3.add_alarm_connection(control_block_plus)
function_block3.add_alarm_connection(control_block_cross)
function_block4.add_input(function_block6)
function_block4.add_input(reward_acc_H)
function_block5.add_input(function_block7)
function_block6.add_input(reward_ph)
function_block7.add_input(control_blockH)
function_block7.add_input(function_block2)
function_block8.add_input(control_blockH)
discretization_block.add_input(function_block1)
computational_graph = ComputationalGraph(blocks=blocks, model=mdp)
core = HierarchicalCore(computational_graph)
# Train
low_level_dataset_eval1 = list()
low_level_dataset_eval2 = list()
dataset_eval = list()
dataset_eval_run = core.evaluate(n_episodes=ep_per_run)
# print('distribution parameters: ', distribution.get_parameters())
J = compute_J(dataset_eval_run, gamma=mdp.info.gamma)
dataset_eval += dataset_eval_run
print('J at start : ' + str(np.mean(J)))
for n in range(n_runs):
print('ITERATION', n)
core.learn(n_episodes=n_iterations * ep_per_run, skip=True)
dataset_eval_run = core.evaluate(n_episodes=ep_per_run)
J = compute_J(dataset_eval_run, gamma=mdp.info.gamma)
print('J at iteration ' + str(n) + ': ' + str(np.mean(J)))
dataset_eval += dataset_eval_run
dataset_plus = control_block_plus.dataset.get()
J_plus = compute_J(dataset_plus, mdp.info.gamma)
dataset_cross = control_block_cross.dataset.get()
J_cross = compute_J(dataset_cross, mdp.info.gamma)
low_level_dataset_eval1.append(dataset_plus)
low_level_dataset_eval2.append(dataset_cross)
print('J ll PLUS at iteration ' + str(n) + ': ' +
str(np.mean(J_plus)))
print('J ll CROSS at iteration ' + str(n) + ': ' +
str(np.mean(J_cross)))
if n == 4:
control_blockH.callbacks = [epsilon_update]
# Tile data
hi_lev_params = agentH.Q.table
max_q_val = np.zeros(n_tiles_high[0]**2)
act_max_q_val = np.zeros(n_tiles_high[0]**2)
for n in range(n_tiles_high[0]**2):
max_q_val[n] = np.amax(hi_lev_params[n])
act_max_q_val[n] = np.argmax(hi_lev_params[n])
mk_dir_recursive('./' + subdir + str(i))
np.save(subdir + str(i) + '/low_level_dataset1_file',
low_level_dataset_eval1)
np.save(subdir + str(i) + '/low_level_dataset2_file',
low_level_dataset_eval2)
np.save(subdir + str(i) + '/max_q_val_tiled_file', max_q_val)
np.save(subdir + str(i) + '/act_max_q_val_tiled_file', act_max_q_val)
np.save(subdir + str(i) + '/dataset_eval_file', dataset_eval)
return
def experiment(n_epochs, n_steps, n_steps_test):
np.random.seed()
# MDP
horizon = 1000
gamma = 0.99
gamma_eval = 1.
mdp = Gym('Acrobot-v1', horizon, gamma)
# Policy
epsilon = LinearDecayParameter(value=1., min_value=.01, n=5000)
epsilon_test = Parameter(value=0.)
epsilon_random = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon_random)
# Settings
initial_replay_size = 500
max_replay_size = 5000
target_update_frequency = 100
batch_size = 200
n_features = 80
train_frequency = 1
# Approximator
input_shape = mdp.info.observation_space.shape
approximator_params = dict(network=Network,
optimizer={
'class': optim.Adam,
'params': {
'lr': .001
}
},
loss=F.smooth_l1_loss,
n_features=n_features,
input_shape=input_shape,
output_shape=mdp.info.action_space.size,
n_actions=mdp.info.action_space.n)
# Agent
agent = DQN(PyTorchApproximator,
pi,
mdp.info,
approximator_params=approximator_params,
batch_size=batch_size,
n_approximators=1,
initial_replay_size=initial_replay_size,
max_replay_size=max_replay_size,
target_update_frequency=target_update_frequency)
# Algorithm
core = Core(agent, mdp)
core.learn(n_steps=initial_replay_size,
n_steps_per_fit=initial_replay_size)
# RUN
pi.set_epsilon(epsilon_test)
dataset = core.evaluate(n_steps=n_steps_test, render=False)
J = compute_J(dataset, gamma_eval)
print('J: ', np.mean(J))
for n in range(n_epochs):
print('Epoch: ', n)
pi.set_epsilon(epsilon)
core.learn(n_steps=n_steps, n_steps_per_fit=train_frequency)
pi.set_epsilon(epsilon_test)
dataset = core.evaluate(n_steps=n_steps_test, render=False)
J = compute_J(dataset, gamma_eval)
print('J: ', np.mean(J))
print('Press a button to visualize acrobot')
input()
core.evaluate(n_episodes=5, render=True)
def experiment():
np.random.seed()
# Argument parser
parser = argparse.ArgumentParser()
arg_game = parser.add_argument_group('Game')
arg_game.add_argument("--name",
type=str,
default='BreakoutDeterministic-v4',
help='Gym ID of the Atari game.')
arg_game.add_argument("--screen-width", type=int, default=84,
help='Width of the game screen.')
arg_game.add_argument("--screen-height", type=int, default=84,
help='Height of the game screen.')
arg_mem = parser.add_argument_group('Replay Memory')
arg_mem.add_argument("--initial-replay-size", type=int, default=50000,
help='Initial size of the replay memory.')
arg_mem.add_argument("--max-replay-size", type=int, default=500000,
help='Max size of the replay memory.')
arg_net = parser.add_argument_group('Deep Q-Network')
arg_net.add_argument("--optimizer",
choices=['adadelta',
'adam',
'rmsprop',
'rmspropcentered'],
default='adam',
help='Name of the optimizer to use.')
arg_net.add_argument("--learning-rate", type=float, default=.00025,
help='Learning rate value of the optimizer.')
arg_net.add_argument("--decay", type=float, default=.95,
help='Discount factor for the history coming from the'
'gradient momentum in rmspropcentered and'
'rmsprop')
arg_net.add_argument("--epsilon", type=float, default=.01,
help='Epsilon term used in rmspropcentered and'
'rmsprop')
arg_alg = parser.add_argument_group('Algorithm')
arg_alg.add_argument("--algorithm", choices=['dqn', 'ddqn', 'adqn'],
default='dqn',
help='Name of the algorithm. dqn is for standard'
'DQN, ddqn is for Double DQN and adqn is for'
'Averaged DQN.')
arg_alg.add_argument("--n-approximators", type=int, default=1,
help="Number of approximators used in the ensemble for"
"Averaged DQN.")
arg_alg.add_argument("--batch-size", type=int, default=32,
help='Batch size for each fit of the network.')
arg_alg.add_argument("--history-length", type=int, default=4,
help='Number of frames composing a state.')
arg_alg.add_argument("--target-update-frequency", type=int, default=10000,
help='Number of collected samples before each update'
'of the target network.')
arg_alg.add_argument("--evaluation-frequency", type=int, default=250000,
help='Number of collected samples before each'
'evaluation. An epoch ends after this number of'
'steps')
arg_alg.add_argument("--train-frequency", type=int, default=4,
help='Number of collected samples before each fit of'
'the neural network.')
arg_alg.add_argument("--max-steps", type=int, default=50000000,
help='Total number of collected samples.')
arg_alg.add_argument("--final-exploration-frame", type=int, default=1000000,
help='Number of collected samples until the exploration'
'rate stops decreasing.')
arg_alg.add_argument("--initial-exploration-rate", type=float, default=1.,
help='Initial value of the exploration rate.')
arg_alg.add_argument("--final-exploration-rate", type=float, default=.1,
help='Final value of the exploration rate. When it'
'reaches this values, it stays constant.')
arg_alg.add_argument("--test-exploration-rate", type=float, default=.05,
help='Exploration rate used during evaluation.')
arg_alg.add_argument("--test-samples", type=int, default=125000,
help='Number of collected samples for each'
'evaluation.')
arg_alg.add_argument("--max-no-op-actions", type=int, default=8,
help='Maximum number of no-op actions performed at the'
'beginning of the episodes. The minimum number is'
'history_length. This number is reported to be 30'
'in the DQN Deepmind paper but, since they'
'consider the first 30 frames without frame'
'skipping and that the number of skipped frames'
'is generally 4, we set it to 8.')
arg_alg.add_argument("--no-op-action-value", type=int, default=0,
help='Value of the no-op action.')
arg_utils = parser.add_argument_group('Utils')
arg_utils.add_argument('--device', type=int, default=None,
help='ID of the GPU device to use. If None, CPU is'
'used.')
arg_utils.add_argument('--load-path', type=str,
help='Path of the model to be loaded.')
arg_utils.add_argument('--save', action='store_true',
help='Flag specifying whether to save the model.')
arg_utils.add_argument('--render', action='store_true',
help='Flag specifying whether to render the game.')
arg_utils.add_argument('--quiet', action='store_true',
help='Flag specifying whether to hide the progress'
'bar.')
arg_utils.add_argument('--debug', action='store_true',
help='Flag specifying whether the script has to be'
'run in debug mode.')
args = parser.parse_args()
scores = list()
optimizer = dict()
if args.optimizer == 'adam':
optimizer['class'] = optim.Adam
optimizer['params'] = dict(lr=args.learning_rate)
elif args.optimizer == 'adadelta':
optimizer['class'] = optim.Adadelta
optimizer['params'] = dict(lr=args.learning_rate)
elif args.optimizer == 'rmsprop':
optimizer['class'] = optim.RMSprop
optimizer['params'] = dict(lr=args.learning_rate,
alpha=args.decay,
eps=args.epsilon)
elif args.optimizer == 'rmspropcentered':
optimizer['class'] = optim.RMSprop
optimizer['params'] = dict(lr=args.learning_rate,
alpha=args.decay,
eps=args.epsilon,
centered=True)
else:
raise ValueError
# Evaluation of the model provided by the user.
if args.load_path:
# MDP
mdp = Atari(args.name, args.screen_width, args.screen_height,
ends_at_life=False)
# Policy
epsilon_test = Parameter(value=args.test_exploration_rate)
pi = EpsGreedy(epsilon=epsilon_test)
# Approximator
input_shape = (args.screen_height, args.screen_width,
args.history_length)
approximator_params = dict(
network=Network,
input_shape=input_shape,
output_shape=(mdp.info.action_space.n,),
n_actions=mdp.info.action_space.n,
load_path=args.load_path,
optimizer=optimizer,
loss=F.smooth_l1_loss,
device=args.device
)
approximator = PyTorchApproximator
# Agent
algorithm_params = dict(
batch_size=1,
train_frequency=1,
target_update_frequency=1,
initial_replay_size=0,
max_replay_size=0,
history_length=args.history_length,
max_no_op_actions=args.max_no_op_actions,
no_op_action_value=args.no_op_action_value,
dtype=np.uint8
)
agent = DQN(approximator, pi, mdp.info,
approximator_params=approximator_params, **algorithm_params)
# Algorithm
core_test = Core(agent, mdp)
# Evaluate model
pi.set_epsilon(epsilon_test)
dataset = core_test.evaluate(n_steps=args.test_samples,
render=args.render,
quiet=args.quiet)
get_stats(dataset)
else:
# DQN learning run
# Summary folder
folder_name = './logs/atari_' + args.algorithm + '_' + args.name +\
'_' + datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
pathlib.Path(folder_name).mkdir(parents=True)
# Settings
if args.debug:
initial_replay_size = 50
max_replay_size = 500
train_frequency = 5
target_update_frequency = 10
test_samples = 20
evaluation_frequency = 50
max_steps = 1000
else:
initial_replay_size = args.initial_replay_size
max_replay_size = args.max_replay_size
train_frequency = args.train_frequency
target_update_frequency = args.target_update_frequency
test_samples = args.test_samples
evaluation_frequency = args.evaluation_frequency
max_steps = args.max_steps
# MDP
mdp = Atari(args.name, args.screen_width, args.screen_height,
ends_at_life=True)
# Policy
epsilon = LinearDecayParameter(value=args.initial_exploration_rate,
min_value=args.final_exploration_rate,
n=args.final_exploration_frame)
epsilon_test = Parameter(value=args.test_exploration_rate)
epsilon_random = Parameter(value=1)
pi = EpsGreedy(epsilon=epsilon_random)
# Approximator
input_shape = (args.screen_height, args.screen_width,
args.history_length)
approximator_params = dict(
network=Network,
input_shape=input_shape,
output_shape=(mdp.info.action_space.n,),
n_actions=mdp.info.action_space.n,
folder_name=folder_name,
optimizer=optimizer,
loss=F.smooth_l1_loss,
device=args.device
)
approximator = PyTorchApproximator
# Agent
algorithm_params = dict(
batch_size=args.batch_size,
n_approximators=args.n_approximators,
initial_replay_size=initial_replay_size,
max_replay_size=max_replay_size,
history_length=args.history_length,
target_update_frequency=target_update_frequency//train_frequency,
max_no_op_actions=args.max_no_op_actions,
no_op_action_value=args.no_op_action_value,
dtype=np.uint8
)
if args.algorithm == 'dqn':
agent = DQN(approximator, pi, mdp.info,
approximator_params=approximator_params,
**algorithm_params)
elif args.algorithm == 'ddqn':
agent = DoubleDQN(approximator, pi, mdp.info,
approximator_params=approximator_params,
**algorithm_params)
elif args.algorithm == 'adqn':
agent = AveragedDQN(approximator, pi, mdp.info,
approximator_params=approximator_params,
**algorithm_params)
# Algorithm
core = Core(agent, mdp)
# RUN
# Fill replay memory with random dataset
print_epoch(0)
core.learn(n_steps=initial_replay_size,
n_steps_per_fit=initial_replay_size, quiet=args.quiet)
if args.save:
agent.approximator.model.save()
# Evaluate initial policy
pi.set_epsilon(epsilon_test)
if args.algorithm == 'ddqn':
agent.policy.set_q(agent.target_approximator)
mdp.set_episode_end(False)
dataset = core.evaluate(n_steps=test_samples, render=args.render,
quiet=args.quiet)
scores.append(get_stats(dataset))
if args.algorithm == 'ddqn':
agent.policy.set_q(agent.approximator)
np.save(folder_name + '/scores.npy', scores)
for n_epoch in range(1, max_steps // evaluation_frequency + 1):
print_epoch(n_epoch)
print('- Learning:')
# learning step
pi.set_epsilon(epsilon)
mdp.set_episode_end(True)
core.learn(n_steps=evaluation_frequency,
n_steps_per_fit=train_frequency, quiet=args.quiet)
if args.save:
agent.approximator.model.save()
print('- Evaluation:')
# evaluation step
pi.set_epsilon(epsilon_test)
if args.algorithm == 'ddqn':
agent.policy.set_q(agent.target_approximator)
mdp.set_episode_end(False)
dataset = core.evaluate(n_steps=test_samples, render=args.render,
quiet=args.quiet)
scores.append(get_stats(dataset))
if args.algorithm == 'ddqn':
agent.policy.set_q(agent.approximator)
np.save(folder_name + '/scores.npy', scores)
return scores
def experiment(n_epochs, n_steps, n_steps_test):
np.random.seed()
# MDP
horizon = 1000
gamma = 0.99
gamma_eval = 1.
mdp = Gym('Acrobot-v1', horizon, gamma)
# Policy
epsilon = LinearDecayParameter(value=1., min_value=.01, n=5000)
epsilon_test = Parameter(value=0.)
epsilon_random = Parameter(value=1.)
pi = EpsGreedy(epsilon=epsilon_random)
# Settings
initial_replay_size = 500
max_replay_size = 5000
target_update_frequency = 100
batch_size = 200
n_features = 80
train_frequency = 1
# Approximator
input_shape = (1,) + mdp.info.observation_space.shape
approximator_params = dict(network=Network,
optimizer={'class': optim.Adam,
'params': {'lr': .001}},
loss=F.smooth_l1_loss, n_features=n_features,
input_shape=input_shape,
output_shape=mdp.info.action_space.size,
n_actions=mdp.info.action_space.n)
# Agent
agent = DQN(PyTorchApproximator, pi, mdp.info,
approximator_params=approximator_params, batch_size=batch_size,
n_approximators=1, initial_replay_size=initial_replay_size,
max_replay_size=max_replay_size, history_length=1,
target_update_frequency=target_update_frequency,
max_no_op_actions=0, no_op_action_value=0, dtype=np.float32)
# Algorithm
core = Core(agent, mdp)
core.learn(n_steps=initial_replay_size, n_steps_per_fit=initial_replay_size)
# RUN
pi.set_epsilon(epsilon_test)
dataset = core.evaluate(n_steps=n_steps_test, render=False)
J = compute_J(dataset, gamma_eval)
print('J: ', np.mean(J))
for n in range(n_epochs):
print('Epoch: ', n)
pi.set_epsilon(epsilon)
core.learn(n_steps=n_steps, n_steps_per_fit=train_frequency)
pi.set_epsilon(epsilon_test)
dataset = core.evaluate(n_steps=n_steps_test, render=False)
J = compute_J(dataset, gamma_eval)
print('J: ', np.mean(J))
print('Press a button to visualize acrobot')
input()
core.evaluate(n_episodes=5, render=True)
import numpy as np
from sklearn.ensemble import ExtraTreesRegressor
from mushroom.algorithms.value import FQI
from mushroom.core import Core
from mushroom.environments import CarOnHill
from mushroom.policy import EpsGreedy
from mushroom.utils.dataset import compute_J
from mushroom.utils.parameters import Parameter
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
agent = FQI(approximator, pi, mdp.info, n_iterations=20,
approximator_params=approximator_params)
core = Core(agent, mdp)
core.learn(n_episodes=1000, n_episodes_per_fit=1000)