def __init__(self, conf, expl_dims, competence_measure, win_size,
competence_mode, k, progress_mode):
RandomInterest.__init__(self, conf, expl_dims)
self.competence_measure = competence_measure
self.win_size = win_size
self.competence_mode = competence_mode
self.dist_max = np.linalg.norm(self.bounds[0, :] - self.bounds[1, :])
self.k = k
self.progress_mode = progress_mode
self.data_xc = Dataset(len(expl_dims), 1)
self.data_sr = Dataset(len(expl_dims), 0)
self.current_progress = 0.
self.current_interest = 0.
def __init__(self, conf, expl_dims, competence_measure, win_size,
competence_mode, k, progress_mode):
InterestModel.__init__(self, expl_dims)
self.bounds = conf.bounds[:, expl_dims]
self.ndims = self.bounds.shape[1]
self.competence_measure = competence_measure
self.win_size = win_size
self.competence_mode = competence_mode
# self.dist_max = np.linalg.norm(self.bounds[0, :] - self.bounds[1, :])
self.dist_max = 2
self.k = k
self.progress_mode = progress_mode
self.data_xc = Dataset(len(expl_dims), 1)
self.data_sr = Dataset(len(expl_dims), 0)
self.current_progress = 1e-5
self.current_interest = 1e-5
def __init__(self, dim_x, dim_y, **kwargs):
"""Create the forward model
@param dim_x the input dimension
@param dim_y the output dimension
"""
self.dim_x = dim_x
self.dim_y = dim_y
self.dataset = Dataset(dim_x, dim_y)
self.conf = kwargs
def __init__(self, dim_x, dim_y, **kwargs):
"""Create the forward model
@param dim_x the input dimension
@param dim_y the output dimension
"""
self.dim_x = dim_x
self.dim_y = dim_y
self.dataset = Dataset(dim_x, dim_y)
self.conf = kwargs
def __init__(self, conf, expl_dims, x_card, cells_win_size, eps_random,
measure, win_size, competence_measure, competence_mode, k,
progress_mode):
DiscretizedProgress.__init__(self, conf, expl_dims, x_card,
cells_win_size, eps_random, measure)
self.bounds = conf.bounds[:, expl_dims]
self.ndims = self.bounds.shape[1]
self.competence_measure = competence_measure
self.win_size = win_size
self.competence_mode = competence_mode
# self.dist_max_comp = np.linalg.norm(self.bounds[0, :] - self.bounds[1, :])
self.dist_max_comp = 2
self.k = k
self.progress_mode = progress_mode
self.data_xc = Dataset(len(expl_dims), 1)
self.data_sr = Dataset(len(expl_dims), 0)
self.current_progress = 1e-5
self.current_interest = 1e-5
def __init__(self,
conf,
expl_dims,
competence_measure,
win_size,
competence_mode,
k,
progress_mode):
RandomInterest.__init__(self, conf, expl_dims)
self.competence_measure = competence_measure
self.win_size = win_size
self.competence_mode = competence_mode
self.dist_max = np.linalg.norm(self.bounds[0,:] - self.bounds[1,:])
self.k = k
self.progress_mode = progress_mode
self.data_xc = Dataset(len(expl_dims), 1)
self.data_sr = Dataset(len(expl_dims), 0)
self.current_progress = 0.
self.current_interest = 0.
def __init__(self, conf, acquisition, exploration_weight, initial_points, environment, optimisation_iterations, exact_feval):
''' :param string acquisition: choose the model of acquisition function between "MPI","EI" and "LCB"
:param scalar exploration_weight: module the exploration in the acquistion (base 2 for LCB and 0.01 for others)
:param scalar initial_points: the number of initial points to give for the bayesian optimisation
:pram Environment environment: environment on which is used the optimisation
:param scalar optimisation_iterations: number of iterations of the optimisation
:param boolean exact_feval: must be False if the environment is noisy
'''
for attr in ['m_dims', 's_dims']:
setattr(self, attr, getattr(conf, attr))
self.dim_x = len(self.m_dims)
self.dim_y = len(self.s_dims)
self.acquisition = acquisition
self.exploration_weight = exploration_weight
self.initial_points = initial_points
self.dataset = Dataset(len(self.m_dims), len(self.s_dims))
self.conf = conf
self.mode = 'explore'
self.environment = environment
self.optimisation_iterations = optimisation_iterations
self.exact_feval = exact_feval
def __init__(self, conf, acquisition, exploration_weight, initial_points,
environment, optimisation_iterations, exact_feval):
''' :param string acquisition: choose the model of acquisition function between "MPI","EI" and "LCB"
:param scalar exploration_weight: module the exploration in the acquistion (base 2 for LCB and 0.01 for others)
:param scalar initial_points: the number of initial points to give for the bayesian optimisation
:pram Environment environment: environment on which is used the optimisation
:param scalar optimisation_iterations: number of iterations of the optimisation
:param boolean exact_feval: must be False if the environment is noisy
'''
for attr in ['m_dims', 's_dims']:
setattr(self, attr, getattr(conf, attr))
self.dim_x = len(self.m_dims)
self.dim_y = len(self.s_dims)
self.acquisition = acquisition
self.exploration_weight = exploration_weight
self.initial_points = initial_points
self.dataset = Dataset(len(self.m_dims), len(self.s_dims))
self.conf = conf
self.mode = 'explore'
self.environment = environment
self.optimisation_iterations = optimisation_iterations
self.exact_feval = exact_feval
class MiscRandomInterest(RandomInterest):
"""
Add some features to the RandomInterest random babbling class.
Allows to query the recent interest in the whole space,
the recent competence on the babbled points in the whole space,
the competence around a given point based on a mean of the knns.
"""
def __init__(self, conf, expl_dims, competence_measure, win_size,
competence_mode, k, progress_mode):
RandomInterest.__init__(self, conf, expl_dims)
self.competence_measure = competence_measure
self.win_size = win_size
self.competence_mode = competence_mode
self.dist_max = np.linalg.norm(self.bounds[0, :] - self.bounds[1, :])
self.k = k
self.progress_mode = progress_mode
self.data_xc = Dataset(len(expl_dims), 1)
self.data_sr = Dataset(len(expl_dims), 0)
self.current_progress = 0.
self.current_interest = 0.
def add_xc(self, x, c):
self.data_xc.add_xy(x, [c])
def add_sr(self, x):
self.data_sr.add_xy(x)
def update_interest(self, i):
self.current_progress += (1. / self.win_size) * (i -
self.current_progress)
self.current_interest = abs(self.current_progress)
def update(self, xy, ms, snnp=None, sp=None):
c = self.competence_measure(xy[self.expl_dims],
ms[self.expl_dims],
dist_max=self.dist_max)
if self.progress_mode == 'local':
interest = self.interest_xc(xy[self.expl_dims], c)
self.update_interest(interest)
elif self.progress_mode == 'global':
pass
self.add_xc(xy[self.expl_dims], c)
self.add_sr(ms[self.expl_dims])
return interest
def n_points(self):
return len(self.data_xc)
def competence_global(self, mode='sw'):
if self.n_points() > 0:
if mode == 'all':
return np.mean(self.data_c)
elif mode == 'sw':
idxs = range(self.n_points())[-self.win_size:]
return np.mean([self.data_xc.get_y(idx) for idx in idxs])
else:
raise NotImplementedError
else:
return 0.
def mean_competence_pt(self, x):
if self.n_points() > self.k:
_, idxs = self.data_xc.nn_x(x, k=self.k)
return np.mean([self.data_xc.get_y(idx) for idx in idxs])
else:
return self.competence()
def interest_xc(self, x, c):
if self.n_points() > 0:
idx_sg_NN = self.data_xc.nn_x(x, k=1)[1][0]
sr_NN = self.data_sr.get_x(idx_sg_NN)
c_old = competence_dist(x, sr_NN, dist_max=self.dist_max)
return c - c_old
#return np.abs(c - c_old)
else:
return 0.
def interest_pt(self, x):
if self.n_points() > self.k:
_, idxs = self.data_xc.nn_x(x, k=self.k)
idxs = sorted(idxs)
v = [self.data_xc.get_y(idx) for idx in idxs]
n = len(v)
comp_beg = np.mean(v[:int(float(n) / 2.)])
comp_end = np.mean(v[int(float(n) / 2.):])
return np.abs(comp_end - comp_beg)
else:
return self.interest_global()
def interest_global(self):
if self.n_points() < 2:
return 0.
else:
idxs = range(self.n_points())[-self.win_size:]
v = [self.data_xc.get_y(idx) for idx in idxs]
n = len(v)
comp_beg = np.mean(v[:int(float(n) / 2.)])
comp_end = np.mean(v[int(float(n) / 2.):])
return np.abs(comp_end - comp_beg)
def competence(self):
return self.competence_global()
def interest(self):
if self.progress_mode == 'local':
return self.current_interest
elif self.progress_mode == 'global':
return self.interest_global()
else:
raise NotImplementedError
class MiscDiscretizedInterest(DiscretizedProgress):
"""
TODO
Add some features to the RandomInterest random babbling class.
TODO
Allows to query the recent interest in the whole space,
the recent competence on the babbled points in the whole space,
the competence around a given point based on a mean of the knns.
"""
def __init__(self, conf, expl_dims, x_card, cells_win_size, eps_random,
measure, win_size, competence_measure, competence_mode, k,
progress_mode):
DiscretizedProgress.__init__(self, conf, expl_dims, x_card,
cells_win_size, eps_random, measure)
self.bounds = conf.bounds[:, expl_dims]
self.ndims = self.bounds.shape[1]
self.competence_measure = competence_measure
self.win_size = win_size
self.competence_mode = competence_mode
# self.dist_max_comp = np.linalg.norm(self.bounds[0, :] - self.bounds[1, :])
self.dist_max_comp = 2
self.k = k
self.progress_mode = progress_mode
self.data_xc = Dataset(len(expl_dims), 1)
self.data_sr = Dataset(len(expl_dims), 0)
self.current_progress = 1e-5
self.current_interest = 1e-5
def add_xc(self, x, c):
self.data_xc.add_xy(x, [c])
def add_sr(self, x):
self.data_sr.add_xy(x)
def update_interest(self, i):
self.current_progress += (1. / self.win_size) * (i -
self.current_progress)
self.current_interest = abs(self.current_progress)
def update(self, xy, ms, snnp=None, sp=None):
# We update the competence in each cell
comp = self.measure(xy,
ms,
dist_min=self.dist_min,
dist_max=self.dist_max)
x = xy[self.expl_dims]
x_index = self.space.index(x)
ms_expl = ms[self.expl_dims]
ms_index = self.space.index(ms_expl)
# Only give competence if observed s is in the same cell as goal x
# to avoid random fluctuations of progress due to random choices in the other cells and not to competence variations
if ms_index == x_index:
self.discrete_progress.update_from_index_and_competence(
x_index, self.normalize_measure(comp))
# Novelty bonus: if novel cell is reached, give it competence (= interest for win_size iterations)
if sum([qi for qi in self.discrete_progress.queues[ms_index]]) == 0.:
self.discrete_progress.update_from_index_and_competence(
ms_index, self.normalize_measure(self.comp_max))
# We track interest of module
c = self.competence_measure(xy[self.expl_dims],
ms[self.expl_dims],
dist_max=self.dist_max_comp)
if self.progress_mode == 'local':
interest = self.interest_xc(xy[self.expl_dims], c)
self.update_interest(interest)
elif self.progress_mode == 'global':
pass
self.add_xc(xy[self.expl_dims], c)
self.add_sr(ms[self.expl_dims])
return interest
def n_points(self):
return len(self.data_xc)
def competence_global(self, mode='sw'):
if self.n_points() > 0:
if mode == 'all':
return np.mean(self.data_c)
elif mode == 'sw':
idxs = range(self.n_points())[-self.win_size:]
return np.mean([self.data_xc.get_y(idx) for idx in idxs])
else:
raise NotImplementedError
else:
return 0.
def mean_competence_pt(self, x):
if self.n_points() > self.k:
_, idxs = self.data_xc.nn_x(x, k=self.k)
return np.mean([self.data_xc.get_y(idx) for idx in idxs])
else:
return self.competence()
def interest_xc(self, x, c):
if self.n_points() > 0:
idx_sg_NN = self.data_xc.nn_x(x, k=1)[1][0]
sr_NN = self.data_sr.get_x(idx_sg_NN)
c_old = self.competence_measure(x,
sr_NN,
dist_max=self.dist_max_comp)
# c_old = competence_dist(x, sr_NN, dist_max=self.dist_max) # Bug ? why use competence_dist ?
return c - c_old
# return np.abs(c - c_old)
else:
return 0.
def interest_pt(self, x):
if self.n_points() > self.k:
_, idxs = self.data_xc.nn_x(x, k=self.k)
idxs = sorted(idxs)
v = [self.data_xc.get_y(idx) for idx in idxs]
n = len(v)
comp_beg = np.mean(v[:int(float(n) / 2.)])
comp_end = np.mean(v[int(float(n) / 2.):])
return np.abs(comp_end - comp_beg)
else:
return self.interest_global()
def interest_global(self):
if self.n_points() < 2:
return 0.
else:
idxs = range(self.n_points())[-self.win_size:]
v = [self.data_xc.get_y(idx) for idx in idxs]
n = len(v)
comp_beg = np.mean(v[:int(float(n) / 2.)])
comp_end = np.mean(v[int(float(n) / 2.):])
return np.abs(comp_end - comp_beg)
def competence(self):
return self.competence_global()
def interest(self):
if self.progress_mode == 'local':
return self.current_interest
elif self.progress_mode == 'global':
return self.interest_global()
else:
raise NotImplementedError
class MiscGaussianInterest(InterestModel):
"""
Add some features to the RandomInterest random babbling class.
Allows to query the recent interest in the whole space,
the recent competence on the babbled points in the whole space,
the competence around a given point based on a mean of the knns.
"""
def __init__(self, conf, expl_dims, competence_measure, win_size,
competence_mode, k, progress_mode):
InterestModel.__init__(self, expl_dims)
self.bounds = conf.bounds[:, expl_dims]
self.ndims = self.bounds.shape[1]
self.competence_measure = competence_measure
self.win_size = win_size
self.competence_mode = competence_mode
# self.dist_max = np.linalg.norm(self.bounds[0, :] - self.bounds[1, :])
self.dist_max = 2
self.k = k
self.progress_mode = progress_mode
self.data_xc = Dataset(len(expl_dims), 1)
self.data_sr = Dataset(len(expl_dims), 0)
self.current_progress = 1e-5
self.current_interest = 1e-5
def sample(self):
return np.clip(np.random.randn(self.ndims),
a_min=self.bounds[0],
a_max=self.bounds[1])
def sample_given_context(self, c, c_dims):
'''
Sample randomly on dimensions not in context
c: context value on c_dims dimensions, not used
c_dims: w.r.t sensori space dimensions
'''
return self.sample()[list(set(range(self.ndims)) - set(c_dims))]
def add_xc(self, x, c):
self.data_xc.add_xy(x, [c])
def add_sr(self, x):
self.data_sr.add_xy(x)
def update_interest(self, i):
self.current_progress += (1. / self.win_size) * (i -
self.current_progress)
self.current_interest = abs(self.current_progress)
def update(self, xy, ms, snnp=None, sp=None):
c = self.competence_measure(xy[self.expl_dims],
ms[self.expl_dims],
dist_max=self.dist_max)
if self.progress_mode == 'local':
interest = self.interest_xc(xy[self.expl_dims], c)
self.update_interest(interest)
elif self.progress_mode == 'global':
pass
self.add_xc(xy[self.expl_dims], c)
self.add_sr(ms[self.expl_dims])
return interest
def n_points(self):
return len(self.data_xc)
def competence_global(self, mode='sw'):
if self.n_points() > 0:
if mode == 'all':
return np.mean(self.data_c)
elif mode == 'sw':
idxs = range(self.n_points())[-self.win_size:]
return np.mean([self.data_xc.get_y(idx) for idx in idxs])
else:
raise NotImplementedError
else:
return 0.
def mean_competence_pt(self, x):
if self.n_points() > self.k:
_, idxs = self.data_xc.nn_x(x, k=self.k)
return np.mean([self.data_xc.get_y(idx) for idx in idxs])
else:
return self.competence()
def interest_xc(self, x, c):
if self.n_points() > 0:
idx_sg_NN = self.data_xc.nn_x(x, k=1)[1][0]
sr_NN = self.data_sr.get_x(idx_sg_NN)
c_old = self.competence_measure(x, sr_NN, dist_max=self.dist_max)
# c_old = competence_dist(x, sr_NN, dist_max=self.dist_max) # Bug ? why use competence_dist ?
return c - c_old
# return np.abs(c - c_old)
else:
return 0.
def interest_pt(self, x):
if self.n_points() > self.k:
_, idxs = self.data_xc.nn_x(x, k=self.k)
idxs = sorted(idxs)
v = [self.data_xc.get_y(idx) for idx in idxs]
n = len(v)
comp_beg = np.mean(v[:int(float(n) / 2.)])
comp_end = np.mean(v[int(float(n) / 2.):])
return np.abs(comp_end - comp_beg)
else:
return self.interest_global()
def interest_global(self):
if self.n_points() < 2:
return 0.
else:
idxs = range(self.n_points())[-self.win_size:]
v = [self.data_xc.get_y(idx) for idx in idxs]
n = len(v)
comp_beg = np.mean(v[:int(float(n) / 2.)])
comp_end = np.mean(v[int(float(n) / 2.):])
return np.abs(comp_end - comp_beg)
def competence(self):
return self.competence_global()
def interest(self):
if self.progress_mode == 'local':
return self.current_interest
elif self.progress_mode == 'global':
return self.interest_global()
else:
raise NotImplementedError
class BayesianOptimisation(SensorimotorModel):
''' Sensorimotor model using Bayesian optimisation to infer inverse prediction'''
def __init__(self, conf, acquisition, exploration_weight, initial_points, environment, optimisation_iterations, exact_feval):
''' :param string acquisition: choose the model of acquisition function between "MPI","EI" and "LCB"
:param scalar exploration_weight: module the exploration in the acquistion (base 2 for LCB and 0.01 for others)
:param scalar initial_points: the number of initial points to give for the bayesian optimisation
:pram Environment environment: environment on which is used the optimisation
:param scalar optimisation_iterations: number of iterations of the optimisation
:param boolean exact_feval: must be False if the environment is noisy
'''
for attr in ['m_dims', 's_dims']:
setattr(self, attr, getattr(conf, attr))
self.dim_x = len(self.m_dims)
self.dim_y = len(self.s_dims)
self.acquisition = acquisition
self.exploration_weight = exploration_weight
self.initial_points = initial_points
self.dataset = Dataset(len(self.m_dims), len(self.s_dims))
self.conf = conf
self.mode = 'explore'
self.environment = environment
self.optimisation_iterations = optimisation_iterations
self.exact_feval = exact_feval
def infer(self, in_dims, out_dims, x):
if in_dims == self.m_dims and out_dims == self.s_dims: # forward
''' For now only return the nearest neighbor of the motor action '''
assert len(x) == self.dim_x, "Wrong dimension for x. Expected %i, got %i" % (self.dim_x, len(x))
# Find the nearest neighbor of the motor action x
_, index = self.dataset.nn_x(x, k=1)
return self.dataset.get_y(index[0])
elif in_dims == self.s_dims and out_dims == self.m_dims: # inverse
if self.mode == 'exploit':
self.acquisition = 0
assert len(x) == self.dim_y, "Wrong dimension for x. Expected %i, got %i" % (self.dim_y, len(x))
# Find the motor action that lead to the nearest neighbor of the sensitive goal
_, index = self.dataset.nn_y(x, k=1)
x0 = self.dataset.get_x(index[0])
# Find the k nearest neighbors of this motor action
_, index = self.dataset.nn_x(x0, k=self.initial_points)
X = []
Y = []
for i in range(len(index)):
X.append(self.dataset.get_x(index[i]))
Y.append(self.dataset.get_y(index[i]))
# Initialisation of the Bayesian optimisation
func = F_to_minimize(np.array(x), self.dim_x, self.environment)
X_init = np.array(X)
Y_init = func.dist_array(Y)
bounds = []
for i in range(self.dim_x):
bounds.append({'name': 'var_'+ str(i), 'type': 'continuous', 'domain': [self.conf.m_mins[i],self.conf.m_maxs[i]]})
space = GPyOpt.Design_space(bounds)
objective = GPyOpt.core.task.SingleObjective(func.f)
model = GPyOpt.models.GPModel(optimize_restarts=5,verbose=False, exact_feval = self.exact_feval)
acquisition_optimizer = GPyOpt.optimization.AcquisitionOptimizer(space)
if self.acquisition == 'EI':
acquisition = GPyOpt.acquisitions.AcquisitionEI(model, space, acquisition_optimizer, jitter = self.exploration_weight)
elif self.acquisition == 'MPI':
acquisition = GPyOpt.acquisitions.AcquisitionMPI(model, space, acquisition_optimizer, jitter = self.exploration_weight)
elif self.acquisition == 'LCB':
acquisition = GPyOpt.acquisitions.AcquisitionLCB(model, space, acquisition_optimizer, exploration_weight = self.exploration_weight)
else:
raise NotImplementedError
evaluator = GPyOpt.core.evaluators.Sequential(acquisition)
bo = GPyOpt.methods.ModularBayesianOptimization(model, space, objective, acquisition, evaluator, X_init = X_init, Y_init = Y_init)
# Run the optimisation, the eps = -np.inf is set to force the optimisation to do the required number of iterations
bo.run_optimization(max_iter = self.optimisation_iterations, eps = -np.inf)
# Update the model woth the list of points explored during the optimisation
self.list_s = []
for (m,s) in func.pointsList:
self.update(m,s)
return bo.x_opt
else:
raise NotImplementedError
def update(self, m, s):
self.dataset.add_xy(m, s)
def forward_prediction(self, m):
""" Compute the expected sensory effect of the motor command m. It is a shortcut for self.infer(self.conf.m_dims, self.conf.s_dims, m)
"""
return self.infer(self.conf.m_dims, self.conf.s_dims, m)
def inverse_prediction(self, s_g):
""" Compute a motor command to reach the sensory goal s_g. It is a shortcut for self.infer(self.conf.s_dims, self.conf.m_dims, s_g)
"""
return self.infer(self.conf.s_dims, self.conf.m_dims, s_g)
class BayesianOptimisation(SensorimotorModel):
''' Sensorimotor model using Bayesian optimisation to infer inverse prediction'''
def __init__(self, conf, acquisition, exploration_weight, initial_points,
environment, optimisation_iterations, exact_feval):
''' :param string acquisition: choose the model of acquisition function between "MPI","EI" and "LCB"
:param scalar exploration_weight: module the exploration in the acquistion (base 2 for LCB and 0.01 for others)
:param scalar initial_points: the number of initial points to give for the bayesian optimisation
:pram Environment environment: environment on which is used the optimisation
:param scalar optimisation_iterations: number of iterations of the optimisation
:param boolean exact_feval: must be False if the environment is noisy
'''
for attr in ['m_dims', 's_dims']:
setattr(self, attr, getattr(conf, attr))
self.dim_x = len(self.m_dims)
self.dim_y = len(self.s_dims)
self.acquisition = acquisition
self.exploration_weight = exploration_weight
self.initial_points = initial_points
self.dataset = Dataset(len(self.m_dims), len(self.s_dims))
self.conf = conf
self.mode = 'explore'
self.environment = environment
self.optimisation_iterations = optimisation_iterations
self.exact_feval = exact_feval
def infer(self, in_dims, out_dims, x):
if in_dims == self.m_dims and out_dims == self.s_dims: # forward
''' For now only return the nearest neighbor of the motor action '''
assert len(
x
) == self.dim_x, "Wrong dimension for x. Expected %i, got %i" % (
self.dim_x, len(x))
# Find the nearest neighbor of the motor action x
_, index = self.dataset.nn_x(x, k=1)
return self.dataset.get_y(index[0])
elif in_dims == self.s_dims and out_dims == self.m_dims: # inverse
if self.mode == 'exploit':
self.acquisition = 0
assert len(
x
) == self.dim_y, "Wrong dimension for x. Expected %i, got %i" % (
self.dim_y, len(x))
# Find the motor action that lead to the nearest neighbor of the sensitive goal
_, index = self.dataset.nn_y(x, k=1)
x0 = self.dataset.get_x(index[0])
# Find the k nearest neighbors of this motor action
_, index = self.dataset.nn_x(x0, k=self.initial_points)
X = []
Y = []
for i in range(len(index)):
X.append(self.dataset.get_x(index[i]))
Y.append(self.dataset.get_y(index[i]))
# Initialisation of the Bayesian optimisation
func = F_to_minimize(np.array(x), self.dim_x, self.environment)
X_init = np.array(X)
Y_init = func.dist_array(Y)
bounds = []
for i in range(self.dim_x):
bounds.append({
'name':
'var_' + str(i),
'type':
'continuous',
'domain': [self.conf.m_mins[i], self.conf.m_maxs[i]]
})
space = GPyOpt.Design_space(bounds)
objective = GPyOpt.core.task.SingleObjective(func.f)
model = GPyOpt.models.GPModel(optimize_restarts=5,
verbose=False,
exact_feval=self.exact_feval)
acquisition_optimizer = GPyOpt.optimization.AcquisitionOptimizer(
space)
if self.acquisition == 'EI':
acquisition = GPyOpt.acquisitions.AcquisitionEI(
model,
space,
acquisition_optimizer,
jitter=self.exploration_weight)
elif self.acquisition == 'MPI':
acquisition = GPyOpt.acquisitions.AcquisitionMPI(
model,
space,
acquisition_optimizer,
jitter=self.exploration_weight)
elif self.acquisition == 'LCB':
acquisition = GPyOpt.acquisitions.AcquisitionLCB(
model,
space,
acquisition_optimizer,
exploration_weight=self.exploration_weight)
else:
raise NotImplementedError
evaluator = GPyOpt.core.evaluators.Sequential(acquisition)
bo = GPyOpt.methods.ModularBayesianOptimization(model,
space,
objective,
acquisition,
evaluator,
X_init=X_init,
Y_init=Y_init)
# Run the optimisation, the eps = -np.inf is set to force the optimisation to do the required number of iterations
bo.run_optimization(max_iter=self.optimisation_iterations,
eps=-np.inf)
# Update the model woth the list of points explored during the optimisation
self.list_s = []
for (m, s) in func.pointsList:
self.update(m, s)
return bo.x_opt
else:
raise NotImplementedError
def update(self, m, s):
self.dataset.add_xy(m, s)
def forward_prediction(self, m):
""" Compute the expected sensory effect of the motor command m. It is a shortcut for self.infer(self.conf.m_dims, self.conf.s_dims, m)
"""
return self.infer(self.conf.m_dims, self.conf.s_dims, m)
def inverse_prediction(self, s_g):
""" Compute a motor command to reach the sensory goal s_g. It is a shortcut for self.infer(self.conf.s_dims, self.conf.m_dims, s_g)
"""
return self.infer(self.conf.s_dims, self.conf.m_dims, s_g)
class MiscRandomInterest(RandomInterest):
"""
Add some features to the RandomInterest random babbling class.
Allows to query the recent interest in the whole space,
the recent competence on the babbled points in the whole space,
the competence around a given point based on a mean of the knns.
"""
def __init__(self,
conf,
expl_dims,
competence_measure,
win_size,
competence_mode,
k,
progress_mode):
RandomInterest.__init__(self, conf, expl_dims)
self.competence_measure = competence_measure
self.win_size = win_size
self.competence_mode = competence_mode
self.dist_max = np.linalg.norm(self.bounds[0,:] - self.bounds[1,:])
self.k = k
self.progress_mode = progress_mode
self.data_xc = Dataset(len(expl_dims), 1)
self.data_sr = Dataset(len(expl_dims), 0)
self.current_progress = 0.
self.current_interest = 0.
def add_xc(self, x, c):
self.data_xc.add_xy(x, [c])
def add_sr(self, x):
self.data_sr.add_xy(x)
def update_interest(self, i):
self.current_progress += (1. / self.win_size) * (i - self.current_progress)
self.current_interest = abs(self.current_progress)
def update(self, xy, ms, snnp=None, sp=None):
c = self.competence_measure(xy[self.expl_dims], ms[self.expl_dims], dist_max=self.dist_max)
if self.progress_mode == 'local':
interest = self.interest_xc(xy[self.expl_dims], c)
self.update_interest(interest)
elif self.progress_mode == 'global':
pass
self.add_xc(xy[self.expl_dims], c)
self.add_sr(ms[self.expl_dims])
return interest
def n_points(self):
return len(self.data_xc)
def competence_global(self, mode='sw'):
if self.n_points() > 0:
if mode == 'all':
return np.mean(self.data_c)
elif mode == 'sw':
idxs = range(self.n_points())[- self.win_size:]
return np.mean([self.data_xc.get_y(idx) for idx in idxs])
else:
raise NotImplementedError
else:
return 0.
def mean_competence_pt(self, x):
if self.n_points() > self.k:
_, idxs = self.data_xc.nn_x(x, k=self.k)
return np.mean([self.data_xc.get_y(idx) for idx in idxs])
else:
return self.competence()
def interest_xc(self, x, c):
if self.n_points() > 0:
idx_sg_NN = self.data_xc.nn_x(x, k=1)[1][0]
sr_NN = self.data_sr.get_x(idx_sg_NN)
c_old = competence_dist(x, sr_NN, dist_max=self.dist_max)
return c - c_old
#return np.abs(c - c_old)
else:
return 0.
def interest_pt(self, x):
if self.n_points() > self.k:
_, idxs = self.data_xc.nn_x(x, k=self.k)
idxs = sorted(idxs)
v = [self.data_xc.get_y(idx) for idx in idxs]
n = len(v)
comp_beg = np.mean(v[:int(float(n)/2.)])
comp_end = np.mean(v[int(float(n)/2.):])
return np.abs(comp_end - comp_beg)
else:
return self.interest_global()
def interest_global(self):
if self.n_points() < 2:
return 0.
else:
idxs = range(self.n_points())[- self.win_size:]
v = [self.data_xc.get_y(idx) for idx in idxs]
n = len(v)
comp_beg = np.mean(v[:int(float(n)/2.)])
comp_end = np.mean(v[int(float(n)/2.):])
return np.abs(comp_end - comp_beg)
def competence(self): return self.competence_global()
def interest(self):
if self.progress_mode == 'local':
return self.current_interest
elif self.progress_mode == 'global':
return self.interest_global()
else:
raise NotImplementedError
class ForwardModel(object):
"""Class describing the ForwardModel interface"""
@classmethod
def from_dataset(cls, dataset, **kwargs):
"""Construct a Nearest Neighbor forward model from an existing dataset."""
m = cls(dataset.dim_x, dataset.dim_y, **kwargs)
m.dataset = dataset
return m
@classmethod
def from_robot(cls, robot, **kwargs):
"""Construct a Nearest Neighbor forward model from an existing dataset."""
m = cls(len(robot.m_feats), len(robot.s_feats), **kwargs)
return m
def __init__(self, dim_x, dim_y, **kwargs):
"""Create the forward model
@param dim_x the input dimension
@param dim_y the output dimension
"""
self.dim_x = dim_x
self.dim_y = dim_y
self.dataset = Dataset(dim_x, dim_y)
self.conf = kwargs
def reset(self):
self.dataset.reset()
def size(self):
return len(self.dataset)
def add_xy(self, x, y):
"""Add an observation to the forward model
@param x an array of float of length dim_in
@param y an array of float of length dim_out
"""
self.dataset.add_xy(x, y)
def add_xy_batch(self, x_list, y_list):
self.dataset.add_xy_batch(x_list, y_list)
def get_x(self, index):
return self.dataset.get_x(index)
def get_y(self, index):
return self.dataset.get_y(index)
def get_xy(self, index):
return self.dataset.get_xy(index)
def predict_y(self, xq, **kwargs):
"""Provide an prediction of xq in the output space
@param xq an array of float of length dim_x
"""
raise NotImplementedError
def config(self):
"""Return a string with the configuration"""
return ", ".join('%s:%s' % (key, value)
for key, value in self.conf.items())
class ForwardModel(object):
"""Class describing the ForwardModel interface"""
@classmethod
def from_dataset(cls, dataset, **kwargs):
"""Construct a Nearest Neighbor forward model from an existing dataset."""
m = cls(dataset.dim_x, dataset.dim_y, **kwargs)
m.dataset = dataset
return m
@classmethod
def from_robot(cls, robot, **kwargs):
"""Construct a Nearest Neighbor forward model from an existing dataset."""
m = cls(len(robot.m_feats), len(robot.s_feats), **kwargs)
return m
def __init__(self, dim_x, dim_y, **kwargs):
"""Create the forward model
@param dim_x the input dimension
@param dim_y the output dimension
"""
self.dim_x = dim_x
self.dim_y = dim_y
self.dataset = Dataset(dim_x, dim_y)
self.conf = kwargs
def reset(self):
self.dataset.reset()
def size(self):
return len(self.dataset)
def add_xy(self, x, y):
"""Add an observation to the forward model
@param x an array of float of length dim_in
@param y an array of float of length dim_out
"""
self.dataset.add_xy(x, y)
def add_xy_batch(self, x_list, y_list): self.dataset.add_xy_batch(x_list, y_list)
def get_x(self, index):
return self.dataset.get_x(index)
def get_y(self, index):
return self.dataset.get_y(index)
def get_xy(self, index):
return self.dataset.get_xy(index)
def predict_y(self, xq, **kwargs):
"""Provide an prediction of xq in the output space
@param xq an array of float of length dim_x
"""
raise NotImplementedError
def config(self):
"""Return a string with the configuration"""
return ", ".join('%s:%s' % (key, value) for key, value in self.conf.items())
