def test_sparse_domain_from_mapping(self):
p_0 = {(4, 3): 0.1, (9, 4): 0.3, (11, -1): 0.2, (44, 44): 0.4}
states, values = domain.from_mapping(p_0)
goal_states = list(p_0)
for i, goal_state in enumerate(goal_states):
assert_array_equal(states[:, i], goal_state)
assert p_0[goal_state] == values[i]
def test_sparse_domain_from_mapping(self):
p_0 = {(4, 3) : 0.1,
(9, 4) : 0.3,
(11, -1) : 0.2,
(44, 44) : 0.4}
states, values = domain.from_mapping(p_0)
goal_states = list(p_0)
for i, goal_state in enumerate(goal_states):
assert_array_equal(states[:, i], goal_state)
assert p_0[goal_state] == values[i]
def _metavariance(p, exponent):
"""
_metavariance(p, exponent) -> alpha
Returns alpha := E[ product_i (X_i - mu_i)**alpha ], where
i ranges over the dimension of the keys of p.
"""
states, probabilities = domain.from_mapping(p)
mu = expectation((states, probabilities))
diffs = (states - numpy.asarray(mu)[:, numpy.newaxis])
if exponent != 1:
diffs **= exponent
product = numpy.multiply.reduce(diffs, axis=0)
alpha = expectation((product, probabilities))
assert alpha.shape == (1, )
return alpha[0]
def expectation(p):
"""
expectation(p) -> mu
Returns the expected value mu, treating the mapping p as a distribution
p : states -> probabilities.
"""
if type(p) is tuple:
assert len(p) == 2
states, probabilities = p
else:
states, probabilities = domain.from_mapping(p)
weighted_states = states * probabilities[numpy.newaxis, :]
mu = numpy.add.reduce(weighted_states, axis=1)
return mu
def _metavariance(p, exponent):
"""
_metavariance(p, exponent) -> alpha
Returns alpha := E[ product_i (X_i - mu_i)**alpha ], where
i ranges over the dimension of the keys of p.
"""
states, probabilities = domain.from_mapping(p)
mu = expectation((states, probabilities))
diffs = (states - numpy.asarray(mu)[:, numpy.newaxis])
if exponent != 1:
diffs **= exponent
product = numpy.multiply.reduce(diffs, axis = 0)
alpha = expectation((product, probabilities))
assert alpha.shape == (1, )
return alpha[0]
def expectation(p):
"""
expectation(p) -> mu
Returns the expected value mu, treating the mapping p as a distribution
p : states -> probabilities.
"""
if type(p) is tuple:
assert len(p) == 2
states, probabilities = p
else:
states, probabilities = domain.from_mapping(p)
weighted_states = states * probabilities[numpy.newaxis, :]
mu = numpy.add.reduce(weighted_states, axis=1)
return mu
def compress(p, epsilon):
"""
compress(p, epsilon) -> compressed epsilon-approximation of p
Returns an approximation of the mapping p, treating p as a distribution
p : states -> probabilities. The returned approximation is *compressed*,
in the sense that it is the approximation with the smallest support, while
the error between p and the approximation is within epsilon (L1 norm).
"""
if not (0.0 <= epsilon <= 1.0):
raise ValueError('epsilon must be within range: 0.0 <= epsilon <= 1.0')
p_compressed = {}
if len(p) > 0:
# create array representation of distribution
states, probabilities = domain.from_mapping(p)
# order entries with respect to increasing probability
order = numpy.argsort(probabilities)
states = states.transpose()[order]
probabilities = probabilities[order]
# discard the largest number of states while keeping the
# corresponding net probability discarded below epsilon
cumulative_probability = numpy.add.accumulate(probabilities)
approximation = (cumulative_probability >= epsilon)
states = states[approximation]
probabilities = probabilities[approximation]
assert len(states) == len(probabilities)
# convert approximation back to a sparse dictionary format
for state, probability in zip(states, probabilities):
p_compressed[tuple(state)] = probability
return p_compressed
def compress(p, epsilon):
"""
compress(p, epsilon) -> compressed epsilon-approximation of p
Returns an approximation of the mapping p, treating p as a distribution
p : states -> probabilities. The returned approximation is *compressed*,
in the sense that it is the approximation with the smallest support, while
the error between p and the approximation is within epsilon (L1 norm).
"""
if not (0.0 <= epsilon <= 1.0):
raise ValueError('epsilon must be within range: 0.0 <= epsilon <= 1.0')
p_compressed = {}
if len(p) > 0:
# create array representation of distribution
states, probabilities = domain.from_mapping(p)
# order entries with respect to increasing probability
order = numpy.argsort(probabilities)
states = states.transpose()[order]
probabilities = probabilities[order]
# discard the largest number of states while keeping the
# corresponding net probability discarded below epsilon
cumulative_probability = numpy.add.accumulate(probabilities)
approximation = (cumulative_probability >= epsilon)
states = states[approximation]
probabilities = probabilities[approximation]
assert len(states) == len(probabilities)
# convert approximation back to a sparse dictionary format
for state, probability in itertools.izip(states, probabilities):
p_compressed[tuple(state)] = probability
return p_compressed
def pack_distribution(self, p_sparse, p_dense=None):
"""
convenience routine to translate a distribution from a dictionary to
a dense array, using this state enumeration
"""
if p_dense is None:
p_dense = numpy.zeros((self.size, ), dtype=numpy.float)
# guard against case where p_sparse is empty
if len(p_sparse) == 0:
return p_dense
p_states, p_values = domain.from_mapping(p_sparse)
# now sort the states, keeping them synchronised with the
# ordering of the values
order = numpy.lexsort(p_states)
p_states = p_states[:, order]
p_values = p_values[order]
p_indices = self.indices(p_states)
p_dense[p_indices] = p_values
return p_dense
def pack_distribution(self, p_sparse, p_dense=None):
"""
convenience routine to translate a distribution from a dictionary to
a dense array, using this state enumeration
"""
if p_dense is None:
p_dense = numpy.zeros((self.size, ), dtype=numpy.float)
# guard against case where p_sparse is empty
if len(p_sparse) == 0:
return p_dense
p_states, p_values = domain.from_mapping(p_sparse)
# now sort the states, keeping them synchronised with the
# ordering of the values
order = numpy.lexsort(p_states)
p_states = p_states[:, order]
p_values = p_values[order]
p_indices = self.indices(p_states)
p_dense[p_indices] = p_values
return p_dense
def map_distribution(f, p, g=None):
"""
map_distribution(f, p [, g]) -> mapping
Returns a copy of the mapping p, with each key replaced by its
image under f. Any duplicate image keys are merged, with the value of
the merged key equal to the sum of the values.
It is expected that f returns tuples or scalars, and behaves in a
reasonable way when given vector state array arguments.
If g is supplied, it is used instead of addition to reduce the values of
duplicate image keys. If given, g must have a reduce method of the form
g.reduce(probabilities) -> reduced_probability
for example, setting g to a numpy ufunc would be fine.
"""
# all this nonsense actually does something fairly straight forward
# see 'map_distribution_simple' for a reference implementation that
# avoids numpy operations
num_items = len(p)
if num_items == 0:
return {}
if g is None:
g = numpy.add
s, v = domain.from_mapping(p)
fs = numpy.asarray(f(s))
# handle case where f returns scalar arguments, say
# this might be a touch flakey
if len(fs.shape) != 2:
fs = fs * numpy.ones((1, s.shape[1]))
# sort image states using lexical ordering on coords, then
# apply same ordering to values
order = numpy.lexsort(fs)
sfs = fs[:, order]
sv = v[order]
# figure out the indices of the first instance of each state
not_equal_adj = numpy.logical_or.reduce(sfs[:, :-1] != sfs[:, 1:])
not_equal_adj = numpy.concatenate(([True], not_equal_adj))
# extract the unique image states under f
usfs = sfs[:, not_equal_adj]
# convert back from arrya representation to iterator of state tuples
unique_image_states = domain.to_iter(usfs)
# determine start and end indices of each equivalence class of
# values in the sorted values array, where values are equivalent if
# they are associated with states that agree under the transform f
class_begin = numpy.nonzero(not_equal_adj)[0]
class_end = numpy.concatenate((class_begin[1:], [num_items]))
# construct the resulting mapped probability distribution
# each image state s maps to the values in its equivalence class,
# reduced by g
p_mapped = {}
for s, i, j in zip(unique_image_states, class_begin, class_end):
p_mapped[s] = g.reduce(sv[i:j])
return p_mapped
def map_distribution(f, p, g=None):
"""
map_distribution(f, p [, g]) -> mapping
Returns a copy of the mapping p, with each key replaced by its
image under f. Any duplicate image keys are merged, with the value of
the merged key equal to the sum of the values.
It is expected that f returns tuples or scalars, and behaves in a
reasonable way when given vector state array arguments.
If g is supplied, it is used instead of addition to reduce the values of
duplicate image keys. If given, g must have a reduce method of the form
g.reduce(probabilities) -> reduced_probability
for example, setting g to a numpy ufunc would be fine.
"""
# all this nonsense actually does something fairly straight forward
# see 'map_distribution_simple' for a reference implementation that
# avoids numpy operations
num_items = len(p)
if num_items == 0:
return {}
if g is None:
g = numpy.add
s, v = domain.from_mapping(p)
fs = numpy.asarray(f(s))
# handle case where f returns scalar arguments, say
# this might be a touch flakey
if len(fs.shape) != 2:
fs = fs*numpy.ones((1, s.shape[1]))
# sort image states using lexical ordering on coords, then
# apply same ordering to values
order = numpy.lexsort(fs)
sfs = fs[:, order]
sv = v[order]
# figure out the indices of the first instance of each state
not_equal_adj = numpy.logical_or.reduce(sfs[:, :-1] != sfs[:, 1:])
not_equal_adj = numpy.concatenate(([True], not_equal_adj))
# extract the unique image states under f
usfs = sfs[:, not_equal_adj]
# convert back from arrya representation to iterator of state tuples
unique_image_states = domain.to_iter(usfs)
# determine start and end indices of each equivalence class of
# values in the sorted values array, where values are equivalent if
# they are associated with states that agree under the transform f
class_begin = numpy.nonzero(not_equal_adj)[0]
class_end = numpy.concatenate((class_begin[1:], [num_items]))
# construct the resulting mapped probability distribution
# each image state s maps to the values in its equivalence class,
# reduced by g
p_mapped = {}
for s, i, j in itertools.izip(unique_image_states, class_begin, class_end):
p_mapped[s] = g.reduce(sv[i:j])
return p_mapped
