Skip to content

Commit

Permalink
Implemented many mean strategy bounds calculations
Browse files Browse the repository at this point in the history
  • Loading branch information
@mike-gimelfarb
mike-gimelfarb committed Nov 22, 2024
1 parent 9aaa299 commit 97647cd
Showing 1 changed file with 44 additions and 30 deletions.
74 changes: 44 additions & 30 deletions pyRDDLGym/core/intervals.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@ class IntervalAnalysisStrategy(Enum):
PERCENTILE = 2
MEAN = 3


class RDDLIntervalAnalysis:

def __init__(
Expand All @@ -43,8 +44,7 @@ def __init__(
:param percentiles: percentiles used to compute bounds when strategy is set to PERCENTILE
'''
self.rddl = rddl
self.logger = logger

self.logger = logger
self.strategy = strategy
self.percentiles = percentiles

Expand Down Expand Up @@ -914,9 +914,8 @@ def _bound_random(self, expr, intervals):
return self._bound_student(expr, intervals)
elif name == 'Gumbel':
return self._bound_gumbel(expr, intervals)
# not implemented
# elif name == 'Laplace':
# return self._bound_laplace(expr, intervals)
elif name == 'Laplace':
return self._bound_laplace(expr, intervals)
elif name == 'Cauchy':
return self._bound_cauchy(expr, intervals)
elif name == 'Gompertz':
Expand All @@ -941,12 +940,7 @@ def _bound_random_kron(self, expr, intervals):
args = expr.args
arg, = args

if self.strategy == IntervalAnalysisStrategy.PERCENTILE:
raise NotImplementedError("Percentile strategy is not implemented for Kronecker delta distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Kronecker delta distribution yet.")

# SUPPORT, PERCENTILE or MEAN strategy
return self._bound(arg, intervals)

def _bound_random_dirac(self, expr, intervals):
Expand All @@ -966,7 +960,9 @@ def _bound_uniform(self, expr, intervals):
raise NotImplementedError("Percentile strategy is not implemented for Uniform distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Uniform distribution yet.")
lower = (la + lb) / 2
upper = (ua + ub) / 2
return (lower, upper)

lower = la
upper = ub
Expand All @@ -979,13 +975,10 @@ def _bound_bernoulli(self, expr, intervals):

if self.strategy == IntervalAnalysisStrategy.PERCENTILE:
lower_percentile, upper_percentile = self.percentiles

lower = np.zeros(shape=np.shape(lp), dtype=np.int64)
upper = np.ones(shape=np.shape(up), dtype=np.int64)

lower = self._mask_assign(lower, lower_percentile > (1 - lp), 1)
upper = self._mask_assign(upper, upper_percentile <= (1 - up), 0)

return (lower, upper)

if self.strategy == IntervalAnalysisStrategy.MEAN:
Expand All @@ -1007,13 +1000,11 @@ def _bound_normal(self, expr, intervals):
if self.strategy == IntervalAnalysisStrategy.PERCENTILE:
# mean + std * normal_inverted_cdf(p)
lower_percentile, upper_percentile = self.percentiles

lower = lm + np.sqrt(lv) * stats.norm.ppf(lower_percentile)
upper = um + np.sqrt(uv) * stats.norm.ppf(upper_percentile)
return (lower, upper)

if self.strategy == IntervalAnalysisStrategy.MEAN:
# mean
return (lm, um)

# SUPPORT strategy
Expand All @@ -1032,7 +1023,7 @@ def _bound_poisson(self, expr, intervals):
raise NotImplementedError("Percentile strategy is not implemented for Poisson distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Poisson distribution yet.")
return (lp, up)

lower = np.zeros(shape=np.shape(lp), dtype=np.int64)
upper = np.full(shape=np.shape(up), fill_value=np.inf, dtype=np.float64)
Expand All @@ -1047,7 +1038,7 @@ def _bound_exponential(self, expr, intervals):
raise NotImplementedError("Percentile strategy is not implemented for Exponential distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Exponential distribution yet.")
return (ls, us)

lower = np.zeros(shape=np.shape(ls), dtype=np.float64)
upper = np.full(shape=np.shape(us), fill_value=np.inf, dtype=np.float64)
Expand All @@ -1062,7 +1053,6 @@ def _bound_weibull(self, expr, intervals):
if self.strategy == IntervalAnalysisStrategy.PERCENTILE:
# scale * (-ln(1 - p))^(1 / shape)
lower_percentile, upper_percentile = self.percentiles

lower = lsc * (-np.log(1 - lower_percentile) ) ** (1 / lsh)
upper = usc * (-np.log(1 - upper_percentile) ) ** (1 / ush)
return (lower, upper)
Expand All @@ -1088,7 +1078,8 @@ def _bound_gamma(self, expr, intervals):
raise NotImplementedError("Percentile strategy is not implemented for Gamma distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Gamma distribution yet.")
# shape * scale
return RDDLIntervalAnalysis._bound_arithmetic_expr((lsh, ush), (lsc, usc), '*')

lower = np.zeros(shape=np.shape(lsh), dtype=np.float64)
upper = np.full(shape=np.shape(usc), fill_value=np.inf, dtype=np.float64)
Expand All @@ -1104,7 +1095,8 @@ def _bound_binomial(self, expr, intervals):
raise NotImplementedError("Percentile strategy is not implemented for Binomial distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Binomial distribution yet.")
# n * p
return RDDLIntervalAnalysis._bound_arithmetic_expr((ln, un), (lp, up), '*')

lower = np.zeros(shape=np.shape(ln), dtype=np.int64)
upper = np.copy(un)
Expand All @@ -1130,15 +1122,16 @@ def _bound_beta(self, expr, intervals):

def _bound_geometric(self, expr, intervals):
args = expr.args
p, = args

p, = args
(lp, up) = self._bound(p, intervals)

if self.strategy == IntervalAnalysisStrategy.PERCENTILE:
raise NotImplementedError("Percentile strategy is not implemented for Geometric distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Geometric distribution yet.")
# 1 / p
one = np.ones_like(up)
return RDDLIntervalAnalysis._bound_arithmetic_expr((one, one), (lp, up), '/')

lower = np.ones(shape=np.shape(lp), dtype=np.int64)
upper = np.full(shape=np.shape(up), fill_value=np.inf, dtype=np.float64)
Expand Down Expand Up @@ -1169,7 +1162,9 @@ def _bound_student(self, expr, intervals):
raise NotImplementedError("Percentile strategy is not implemented for Student distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Student distribution yet.")
lower = np.zeros_like(ld)
upper = np.zeros_like(ud)
return (lower, upper)

lower = np.full(shape=np.shape(ld), fill_value=-np.inf, dtype=np.float64)
upper = np.full(shape=np.shape(ud), fill_value=+np.inf, dtype=np.float64)
Expand All @@ -1185,7 +1180,25 @@ def _bound_gumbel(self, expr, intervals):
raise NotImplementedError("Percentile strategy is not implemented for Gumbel distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Gumbel distribution yet.")
lower = lm + 0.577215664901532 * ls
upper = um + 0.577215664901532 * us
return (lower, upper)

lower = np.full(shape=np.shape(lm), fill_value=-np.inf, dtype=np.float64)
upper = np.full(shape=np.shape(um), fill_value=+np.inf, dtype=np.float64)
return (lower, upper)

def _bound_laplace(self, expr, intervals):
args = expr.args
mean, scale = args
(lm, um) = self._bound(mean, intervals)
(ls, us) = self._bound(scale, intervals)

if self.strategy == IntervalAnalysisStrategy.PERCENTILE:
raise NotImplementedError("Percentile strategy is not implemented for Laplace distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
return (lm, um)

lower = np.full(shape=np.shape(lm), fill_value=-np.inf, dtype=np.float64)
upper = np.full(shape=np.shape(um), fill_value=+np.inf, dtype=np.float64)
Expand All @@ -1201,7 +1214,7 @@ def _bound_cauchy(self, expr, intervals):
raise NotImplementedError("Percentile strategy is not implemented for Cauchy distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Cauchy distribution yet.")
raise ValueError("The mean of a Cauchy distribution is not defined.")

lower = np.full(shape=np.shape(lm), fill_value=-np.inf, dtype=np.float64)
upper = np.full(shape=np.shape(um), fill_value=+np.inf, dtype=np.float64)
Expand Down Expand Up @@ -1232,7 +1245,7 @@ def _bound_chisquare(self, expr, intervals):
raise NotImplementedError("Percentile strategy is not implemented for Chi-square distribution yet.")

if self.strategy == IntervalAnalysisStrategy.MEAN:
raise NotImplementedError("Mean strategy is not implemented for Chi-square distribution yet.")
return (ld, ud)

lower = np.zeros(shape=np.shape(ld), dtype=np.float64)
upper = np.full(shape=np.shape(ud), fill_value=np.inf, dtype=np.float64)
Expand Down Expand Up @@ -1278,4 +1291,5 @@ def _bound_discrete_pvar(self, expr, intervals, unnorm):
lower_prob, upper_prob = self._bound(arg, intervals)
bounds = [(lower_prob[..., i], upper_prob[..., i])
for i in range(lower_prob.shape[-1])]
return self._bound_discrete_helper(bounds)
return self._bound_discrete_helper(bounds)

0 comments on commit 97647cd

Please sign in to comment.