mtl-aas/robustness.py
Marcell Vazquez-Chanlatte aec3c0d15b implemented robustness
2016-10-09 23:14:20 -07:00

108 lines
2.7 KiB
Python

from functools import singledispatch
from operator import sub, add
from lenses import lens
import stl.ast
from stl.utils import set_params, param_lens
oo = float('inf')
@singledispatch
def pointwise_robustness(stl):
raise NotImplementedError
@pointwise_robustness.register(stl.Or)
def _(stl):
return lambda x, t: max(pointwise_robustness(arg)(x, t) for arg in stl.args)
@pointwise_robustness.register(stl.And)
def _(stl):
return lambda x, t: min(pointwise_robustness(arg)(x, t) for arg in stl.args)
@pointwise_robustness.register(stl.F)
def _(stl):
lo, hi = stl.interval
return lambda x, t: max((pointwise_robustness(stl.arg)(x, t + t2)
for t2 in x[lo:hi].index), default=-oo)
@pointwise_robustness.register(stl.G)
def _(stl):
lo, hi = stl.interval
return lambda x, t: min((pointwise_robustness(stl.arg)(x, t + t2)
for t2 in x[lo:hi].index), default=oo)
@pointwise_robustness.register(stl.Neg)
def _(stl):
return lambda x, t: -pointwise_robustness(arg)(x, t)
op_lookup = {
">": sub,
">=": sub,
"<": add,
"<=": add,
"=": lambda a, b: -abs(a - b),
}
@pointwise_robustness.register(stl.LinEq)
def _(stl):
op = op_lookup[stl.op]
return lambda x, t: op(eval_terms(stl, x, t), stl.const)
def eval_terms(lineq, x, t):
psi = lens(lineq).terms.each_().modify(eval_term(x, t))
return sum(psi.terms)
def eval_term(x, t):
return lambda term: term.coeff*x[term.id.name][t]
def binsearch(stleval, *, tol=1e-3, lo, hi, polarity):
"""Only run search if tightest robustness was positive."""
# Only check low since hi taken care of by precondition.
# TODO: allow for different polarities
rL, rH = stleval(lo), stleval(hi)
# Early termination via bounds checks
posL, posH = rL > 0, rH > 0
if polarity and posL:
return lo
elif not polarity and posH:
return hi
while hi - lo > tol:
mid = lo + (hi - lo) / 2
r = stleval(mid)
if not polarity: # swap direction
r *= -1
if r < 0:
lo, hi = mid, hi
else:
lo, hi = lo, mid
# Want satisifiable formula
return hi if polarity else lo
def lex_param_project(stl, x, *, order, polarity, ranges, tol=1e-3):
val = {var: (ranges[var][0] if not polarity[var] else ranges[var][1]) for var in order}
p_lens = param_lens(stl)
def stleval_fact(var, val):
l = lens(val)[var]
return lambda p: pointwise_robustness(set_params(stl, l.set(p)))(x, 0)
for var in order:
stleval = stleval_fact(var, val)
lo, hi = ranges[var]
param = binsearch(stleval, lo=lo, hi=hi, tol=tol, polarity=polarity[var])
val[var] = param
return val