def __init__(self, p, rew, mu=None, gamma=.9, horizon=np.inf):
"""
Constructor.
Args:
p (np.ndarray): transition probability matrix;
rew (np.ndarray): reward matrix;
mu (np.ndarray, None): initial state probability distribution;
gamma (float, .9): discount factor;
horizon (int, np.inf): the horizon.
"""
assert p.shape == rew.shape
assert mu is None or p.shape[0] == mu.size
# MDP parameters
self.p = p
self.r = rew
self.mu = mu
# MDP properties
observation_space = spaces.Discrete(p.shape[0])
action_space = spaces.Discrete(p.shape[1])
horizon = horizon
gamma = gamma
mdp_info = MDPInfo(observation_space, action_space, gamma, horizon)
super().__init__(mdp_info)
def experiment():
np.random.seed(3)
# MDP
mdp = generate_simple_chain(state_n=5,
goal_states=[2],
prob=.8,
rew=1,
gamma=.9)
action_space = mdp._mdp_info.action_space
observation_space = mdp._mdp_info.observation_space
gamma = mdp._mdp_info.gamma
# Model Block
model_block = MBlock(env=mdp, render=False)
#Policy
epsilon = Parameter(value=1)
pi = EpsGreedy(epsilon=epsilon)
table = Table(mdp.info.size)
pi.set_q(table)
#Agents
mdp_info_agent1 = MDPInfo(observation_space=observation_space,
action_space=spaces.Discrete(5),
gamma=1,
horizon=20)
mdp_info_agent2 = MDPInfo(observation_space=spaces.Discrete(5),
action_space=action_space,
gamma=gamma,
horizon=10)
agent1 = SimpleAgent(name='HIGH', mdp_info=mdp_info_agent1, policy=pi)
agent2 = SimpleAgent(name='LOW', mdp_info=mdp_info_agent2, policy=pi)
# Control Blocks
control_block1 = ControlBlock(wake_time=10,
agent=agent1,
n_eps_per_fit=None,
n_steps_per_fit=1)
control_block2 = ControlBlock(wake_time=1,
agent=agent2,
n_eps_per_fit=None,
n_steps_per_fit=1)
# Algorithm
blocks = [model_block, control_block1, control_block2]
order = [0, 1, 2]
model_block.add_input(control_block2)
control_block1.add_input(model_block)
control_block1.add_reward(model_block)
control_block2.add_input(control_block1)
control_block2.add_reward(model_block)
computational_graph = ComputationalGraph(blocks=blocks, order=order)
core = HierarchicalCore(computational_graph)
# Train
core.learn(n_steps=40, quiet=True)
return
def __init__(self, height, width, goal, start=(0, 0)):
# MDP properties
observation_space = spaces.Discrete(height * width)
action_space = spaces.Discrete(4)
horizon = 100
gamma = .9
mdp_info = MDPInfo(observation_space, action_space, gamma, horizon)
super().__init__(mdp_info, height, width, start, goal)
def __init__(self, height=3, width=3, goal=(0, 2), start=(2, 0)):
# MDP properties
observation_space = spaces.Discrete(height * width)
action_space = spaces.Discrete(4)
horizon = np.inf
gamma = .95
mdp_info = MDPInfo(observation_space, action_space, gamma, horizon)
super(GridWorldVanHasselt, self).__init__(mdp_info, height, width,
start, goal)
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 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 __init__(self, grid_map):
self.__name__ = 'GridWorldGenerator'
self._grid, height, width, start, goal = self._generate(grid_map)
# MDP properties
observation_space = spaces.Discrete(height * width)
action_space = spaces.Discrete(4)
horizon = 100
gamma = .9
mdp_info = MDPInfo(observation_space, action_space, gamma, horizon)
super(GridWorldGenerator, self).__init__(mdp_info, height, width,
start, goal)
def __init__(self, grid_map_file, height_window=84, width_window=84):
self.__name__ = 'GridWorldPixelGenerator'
self.window_size = (width_window, height_window)
self._symbols = {
'.': 0.,
'S': 63.75,
'*': 127.5,
'#': 191.25,
'G': 255.
}
self._grid, start, goal = self._generate(grid_map_file)
self._initial_grid = deepcopy(self._grid)
height = self._grid.shape[0]
width = self._grid.shape[1]
assert height_window % height == 0 and width_window % width == 0
# MDP properties
observation_space = spaces.Box(low=0.,
high=255.,
shape=(self.window_size[1],
self.window_size[0]))
action_space = spaces.Discrete(5)
horizon = 100
gamma = .9
mdp_info = MDPInfo(observation_space, action_space, gamma, horizon)
super(GridWorldPixelGenerator, self).__init__(mdp_info, height, width,
start, goal)
def build_discretized_agent(alg, params, n, optim, loss, mdp, eps, n_features,
use_cuda):
high = mdp.info.observation_space.high
low = mdp.info.observation_space.low
observation_space = spaces.Box(low=low, high=high)
action_space = spaces.Discrete(n)
mdp_info = MDPInfo(observation_space=observation_space,
action_space=action_space,
gamma=mdp.info.gamma,
horizon=mdp.info.horizon)
pi = Boltzmann(eps)
approximator_params = dict(network=Network,
optimizer=optim,
loss=loss,
n_features=n_features,
input_shape=mdp_info.observation_space.shape,
output_shape=mdp_info.action_space.size,
n_actions=mdp_info.action_space.n,
use_cuda=use_cuda)
agent = alg(PyTorchApproximator,
pi,
mdp_info,
approximator_params=approximator_params,
**params)
return agent
def __init__(self, p, rew, mu=None, gamma=.9):
self.__name__ = 'FiniteMDP'
assert p.shape == rew.shape
assert mu is None or p.shape[0] == mu.size
# MDP parameters
self.p = p
self.r = rew
self.mu = mu
# MDP properties
observation_space = spaces.Discrete(p.shape[0])
action_space = spaces.Discrete(p.shape[1])
horizon = np.inf
gamma = gamma
mdp_info = MDPInfo(observation_space, action_space, gamma, horizon)
super(FiniteMDP, self).__init__(mdp_info)
def __init__(self,
m=2.,
M=8.,
l=.5,
g=9.8,
mu=1e-2,
max_u=50.,
noise_u=10.,
horizon=3000,
gamma=.95):
"""
Constructor.
Args:
m (float, 2.0): mass of the pendulum;
M (float, 8.0): mass of the cart;
l (float, .5): length of the pendulum;
g (float, 9.8): gravity acceleration constant;
mu (float, 1e-2): friction constant of the pendulum;
max_u (float, 50.): maximum allowed input torque;
noise_u (float, 10.): maximum noise on the action;
horizon (int, 3000): horizon of the problem;
gamma (int, .95): discount factor.
"""
# MDP parameters
self._m = m
self._M = M
self._l = l
self._g = g
self._alpha = 1 / (self._m + self._M)
self._mu = mu
self._dt = .1
self._max_u = max_u
self._noise_u = noise_u
high = np.array([np.inf, np.inf])
# MDP properties
observation_space = spaces.Box(low=-high, high=high)
action_space = spaces.Discrete(3)
mdp_info = MDPInfo(observation_space, action_space, gamma, horizon)
# Visualization
self._viewer = Viewer(2.5 * l, 2.5 * l)
self._last_u = None
self._state = None
super().__init__(mdp_info)
def __init__(self):
self.__name__ = 'CarOnHill'
# MDP parameters
self.max_pos = 1.
self.max_velocity = 3.
high = np.array([self.max_pos, self.max_velocity])
self._g = 9.81
self._m = 1
self._dt = .1
self._discrete_actions = [-4., 4.]
# MDP properties
observation_space = spaces.Box(low=-high, high=high)
action_space = spaces.Discrete(2)
horizon = 100
gamma = .95
mdp_info = MDPInfo(observation_space, action_space, gamma, horizon)
super(CarOnHill, self).__init__(mdp_info)
def __init__(self, horizon=100, gamma=.95):
"""
Constructor.
"""
# MDP parameters
self.max_pos = 1.
self.max_velocity = 3.
high = np.array([self.max_pos, self.max_velocity])
self._g = 9.81
self._m = 1.
self._dt = .1
self._discrete_actions = [-4., 4.]
# MDP properties
observation_space = spaces.Box(low=-high, high=high)
action_space = spaces.Discrete(2)
mdp_info = MDPInfo(observation_space, action_space, gamma, horizon)
super().__init__(mdp_info)
def build_high_level_agent(alg, params, optim, loss, mdp, horizon_low, eps,
n_features, use_cuda):
high = np.ones(4)
low = np.zeros(4)
high[:2] = mdp.info.observation_space.high[:2]
low[:2] = mdp.info.observation_space.low[:2]
high[2:] = mdp.info.observation_space.high[3:5]
low[2:] = mdp.info.observation_space.low[3:5]
n_actions = 9
observation_space = spaces.Box(low=low, high=high)
action_space = spaces.Discrete(n_actions)
mdp_info = MDPInfo(observation_space=observation_space,
action_space=action_space,
gamma=mdp.info.gamma**horizon_low,
horizon=mdp.info.horizon)
pi = Boltzmann(eps)
approximator_params = dict(network=Network,
optimizer=optim,
loss=loss,
n_features=n_features,
input_shape=mdp_info.observation_space.shape,
output_shape=mdp_info.action_space.size,
n_actions=mdp_info.action_space.n,
use_cuda=use_cuda)
agent = alg(PyTorchApproximator,
pi,
mdp_info,
approximator_params=approximator_params,
**params)
return agent
def __init__(self):
self.__name__ = 'InvertedPendulum'
# MDP parameters
self.max_degree = np.inf
self.max_angular_velocity = np.inf
high = np.array([self.max_degree, self.max_angular_velocity])
self._g = 9.8
self._m = 2.
self._M = 8.
self._l = .5
self._alpha = 1. / (self._m + self._M)
self._dt = .1
self._discrete_actions = [-50., 0., 50.]
# MDP properties
observation_space = spaces.Box(low=-high, high=high)
action_space = spaces.Discrete(3)
horizon = 3000
gamma = .95
mdp_info = MDPInfo(observation_space, action_space, gamma, horizon)
super(InvertedPendulum, self).__init__(mdp_info)
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 server_experiment_small(alg_high, alg_low, params, subdir, i):
np.random.seed()
# Model Block
mdp = ShipSteering(small=True, n_steps_action=3)
#State Placeholder
state_ph = PlaceHolder(name='state_ph')
#Reward Placeholder
reward_ph = PlaceHolder(name='reward_ph')
#Last_In Placeholder
lastaction_ph = PlaceHolder(name='lastaction_ph')
# Function Block 1
function_block1 = fBlock(name='f1 (direction to angle difference)',
phi=direction_to_angle)
# Function Block 2
function_block2 = fBlock(name='f2 (cost cosine)', phi=cost_cosine)
#Features
features = Features(basis_list=[PolynomialBasis()])
# Policy 1
epsilon = LinearDecayParameter(value=0.1, min_value=0.0, n=10000)
pi1 = EpsGreedy(epsilon=epsilon)
# Agent 1
learning_rate1 = params.get('learning_rate_high')
lambda_coeff = params.get('lambda_coeff')
mdp_info_agent1 = MDPInfo(observation_space=mdp.info.observation_space,
action_space=spaces.Discrete(8),
gamma=mdp.info.gamma,
horizon=100)
approximator_params1 = dict(input_shape=(features.size, ),
output_shape=mdp_info_agent1.action_space.size,
n_actions=mdp_info_agent1.action_space.n)
agent1 = alg_high(policy=pi1,
mdp_info=mdp_info_agent1,
learning_rate=learning_rate1,
lambda_coeff=lambda_coeff,
features=features,
approximator_params=approximator_params1)
# Control Block 1
control_block1 = ControlBlock(name='Control Block 1',
agent=agent1,
n_steps_per_fit=1)
# Policy 2
pi2 = DeterministicControlPolicy(weights=np.array([0]))
mu2 = np.zeros(pi2.weights_size)
sigma2 = 1e-3 * np.ones(pi2.weights_size)
distribution2 = GaussianDiagonalDistribution(mu2, sigma2)
# Agent 2
learning_rate2 = params.get('learning_rate_low')
mdp_info_agent2 = MDPInfo(observation_space=spaces.Box(
-np.pi, np.pi, (1, )),
action_space=mdp.info.action_space,
gamma=mdp.info.gamma,
horizon=100)
agent2 = alg_low(distribution=distribution2,
policy=pi2,
mdp_info=mdp_info_agent2,
learning_rate=learning_rate2)
# Control Block 2
parameter_callback2 = CollectDistributionParameter(distribution2)
control_block2 = ControlBlock(name='Control Block 2',
agent=agent2,
n_eps_per_fit=10,
callbacks=[parameter_callback2])
#Reward Accumulator
reward_acc = reward_accumulator_block(gamma=mdp_info_agent1.gamma,
name='reward_acc')
# Algorithm
blocks = [
state_ph, reward_ph, lastaction_ph, control_block1, control_block2,
function_block1, function_block2, reward_acc
]
state_ph.add_input(control_block2)
reward_ph.add_input(control_block2)
lastaction_ph.add_input(control_block2)
control_block1.add_input(state_ph)
reward_acc.add_input(reward_ph)
reward_acc.add_alarm_connection(control_block2)
control_block1.add_reward(reward_acc)
control_block1.add_alarm_connection(control_block2)
function_block1.add_input(control_block1)
function_block1.add_input(state_ph)
function_block2.add_input(function_block1)
control_block2.add_input(function_block1)
control_block2.add_reward(function_block2)
computational_graph = ComputationalGraph(blocks=blocks, model=mdp)
core = HierarchicalCore(computational_graph)
# Train
low_level_dataset_eval = list()
dataset_eval = list()
dataset_eval_run = core.evaluate(n_episodes=ep_per_run)
J = compute_J(dataset_eval_run, gamma=mdp.info.gamma)
print('J at start : ' + str(np.mean(J)))
dataset_eval += dataset_eval_run
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)
dataset_eval += dataset_eval_run
J = compute_J(dataset_eval_run, gamma=mdp.info.gamma)
print('J at iteration ' + str(n) + ': ' + str(np.mean(J)))
low_level_dataset_eval += control_block2.dataset.get()
# Save
parameter_dataset2 = parameter_callback2.get_values()
mk_dir_recursive('./' + subdir + str(i))
np.save(subdir + str(i) + '/low_level_dataset_file',
low_level_dataset_eval)
np.save(subdir + str(i) + '/parameter_dataset2_file', parameter_dataset2)
np.save(subdir + str(i) + '/dataset_eval_file', dataset_eval)
return
def experiment():
small = True
print('ENV IS SMALL? ', small)
np.random.seed()
# Model Block
mdp = ShipSteering(small=small, hard=True, 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
lim = 150 if small else 1000
tilingsH = Tiles.generate(n_tilings=1,
n_tiles=[5, 5],
low=[0, 0],
high=[lim, lim])
featuresH = Features(tilings=tilingsH)
# PolicyH
epsilon = LinearDecayParameter(value=0.1, min_value=0.0, n=10000)
piH = EpsGreedy(epsilon=epsilon)
# AgentH
learning_rate = Parameter(value=1)
mdp_info_agentH = MDPInfo(observation_space=spaces.Box(
low=np.array([0, 0]), high=np.array([lim, lim]), shape=(2, )),
action_space=spaces.Discrete(8),
gamma=1,
horizon=10000)
approximator_paramsH = dict(input_shape=(featuresH.size, ),
output_shape=mdp_info_agentH.action_space.size,
n_actions=mdp_info_agentH.action_space.n)
agentH = TrueOnlineSARSALambda(policy=piH,
mdp_info=mdp_info_agentH,
learning_rate=learning_rate,
lambda_coeff=0.9,
approximator_params=approximator_paramsH,
features=featuresH)
# Control Block H
control_blockH = ControlBlock(name='control block H',
agent=agentH,
n_steps_per_fit=1)
#FeaturesL
featuresL = Features(basis_list=[PolynomialBasis()])
# Policy1
input_shape = (featuresL.size, )
approximator_params = dict(input_dim=input_shape[0])
approximator = Regressor(LinearApproximator,
input_shape=input_shape,
output_shape=mdp.info.action_space.shape,
**approximator_params)
sigma = np.array([[1.3e-2]])
pi1 = GaussianPolicy(mu=approximator, sigma=sigma)
# Policy2
pi2 = GaussianPolicy(mu=approximator, sigma=sigma)
# Agent1
learning_rate1 = AdaptiveParameter(value=1e-5)
agent1 = GPOMDP(pi1, mdp.info, learning_rate1, featuresL)
# Agent2
learning_rate2 = AdaptiveParameter(value=1e-5)
agent2 = GPOMDP(pi2, mdp.info, learning_rate2, featuresL)
#Termination Conds
termination_condition1 = TerminationCondition(active_dir=1, small=small)
termination_condition2 = TerminationCondition(active_dir=5, small=small)
# Control Block +
control_block1 = ControlBlock(name='control block 1',
agent=agent1,
n_eps_per_fit=50,
termination_condition=termination_condition1)
# Control Block x
control_block2 = ControlBlock(name='control block 2',
agent=agent2,
n_eps_per_fit=50,
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(small=small), 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(small=small), name='f7 G_lo')
#Reward Accumulator H:
reward_acc_H = reward_accumulator_block(gamma=mdp_info_agentH.gamma,
name='reward_acc_H')
#Mux_Block
mux_block = MuxBlock(name='mux')
mux_block.add_block_list([control_block1])
mux_block.add_block_list([control_block2])
#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, reward_acc_H
]
#state_ph.add_input(mux_block)
#reward_ph.add_input(mux_block)
#lastaction_ph.add_input(mux_block)
reward_acc_H.add_input(reward_ph)
reward_acc_H.add_alarm_connection(control_block1)
reward_acc_H.add_alarm_connection(control_block2)
control_blockH.add_input(function_block1)
control_blockH.add_reward(function_block4)
control_blockH.add_alarm_connection(control_block1)
control_blockH.add_alarm_connection(control_block2)
mux_block.add_input(control_blockH)
mux_block.add_input(function_block2)
control_block1.add_reward(function_block5)
control_block2.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_block1)
function_block3.add_alarm_connection(control_block2)
function_block4.add_input(function_block6)
function_block4.add_input(reward_acc_H)
function_block5.add_input(reward_ph)
function_block5.add_input(function_block7)
function_block6.add_input(reward_ph)
function_block7.add_input(control_blockH)
function_block7.add_input(function_block2)
computational_graph = ComputationalGraph(blocks=blocks, model=mdp)
core = HierarchicalCore(computational_graph)
# Train
dataset_eval_visual = list()
low_level_dataset_eval1 = list()
low_level_dataset_eval2 = list()
n_runs = 5
for n in range(n_runs):
print('ITERATION', n)
core.learn(n_episodes=1000, skip=True)
dataset_eval = core.evaluate(n_episodes=10)
last_ep_dataset = pick_last_ep(dataset_eval)
dataset_eval_visual += last_ep_dataset
low_level_dataset_eval1 += control_block1.dataset.get()
low_level_dataset_eval2 += control_block2.dataset.get()
# Visualize
hi_lev_params = agentH.Q.get_weights()
hi_lev_params = np.reshape(hi_lev_params, (8, 25))
max_q_val = np.zeros(shape=(25, ))
act_max_q_val = np.zeros(shape=(25, ))
for i in range(25):
max_q_val[i] = np.amax(hi_lev_params[:, i])
act_max_q_val[i] = np.argmax(hi_lev_params[:, i])
max_q_val_tiled = np.reshape(max_q_val, (5, 5))
act_max_q_val_tiled = np.reshape(act_max_q_val, (5, 5))
#low_level_dataset1 = dataset_callback1.get()
#low_level_dataset2 = dataset_callback2.get()
subdir = datetime.datetime.now().strftime('%Y-%m-%d_%H-%M-%S') + '/'
mk_dir_recursive('./' + subdir)
np.save(subdir + '/low_level_dataset1_file', low_level_dataset_eval1)
np.save(subdir + '/low_level_dataset2_file', low_level_dataset_eval2)
np.save(subdir + '/max_q_val_tiled_file', max_q_val_tiled)
np.save(subdir + '/act_max_q_val_tiled_file', act_max_q_val_tiled)
np.save(subdir + '/dataset_eval_file', dataset_eval_visual)
return
