started smooth robustness implementation

This commit is contained in:
Marcell Vazquez-Chanlatte 2016-12-10 10:47:35 -08:00
parent 5fc2145622
commit 100f48a0ba

View file

@ -5,59 +5,51 @@ from operator import sub, add
import sympy as sym
from lenses import lens
from numpy import arange
from funcy import pairwise, autocurry
import stl.ast
from stl.ast import t_sym
@singledispatch
def smooth_robustness(stl):
def smooth_robustness(stl, L, h):
raise NotImplementedError
def f1(rs):
@smooth_robustness.register(stl.And)
@smooth_robustness.register(stl.G)
def _(stl, L, H):
raise NotImplementedError("Call canonicalization function")
def soft_max(rs):
return sym.log(sum(sym.exp(r) for r in rs))
def f2(rs):
return sym.log(sum(sym.exp(r) for r in rs)/(len(rs)))
def LSE(rs):
return soft_max(rs) - sym.log(len(rs))
@smooth_robustness.register(stl.Or)
def _(stl, depth=0):
rl, rh = list(zip(*[smooth_robustness(arg, depth) for arg in stl.args]))
return f1(rl), f2(rh)
@smooth_robustness.register(stl.And)
def _(stl, depth=0):
rh, rl = list(zip(*[-smooth_robustness(arg, depth) for arg in stl.args]))
return -f2(rh), -f1(rl)
def _(stl, L, h):
rl, rh = list(zip(
*[smooth_robustness(arg, depth) for arg in stl.args]))
return soft_max(rl), LSE(rh)
def F1(r, interval, t):
lo, hi = interval
bounds = (t, lo, hi)
return sym.log(sym.Integral(sym.exp(r), bounds))
def F2(r, interval, t):
lo, hi = interval
return F1(r, interval, t) - sym.log(hi - lo)
@autocurry
def x_ij(L, h, x_i, x_j):
return (L*h + x_i + x_j)/2
@smooth_robustness.register(stl.F)
def _(stl, depth=0):
depth += 1
t = sym.Symbol("t{}".format(depth))
def _(stl, L, H):
lo, hi = stl.interval
times = arange(lo, hi, H)
rl, rh = smooth_robustness(stl.arg)
r = (rl.subs({t_sym: t}), rh.subs({t_sym: t}))
return F1(r[0], stl.interval, t), F2(rh[1], stl.interval, t)
@smooth_robustness.register(stl.G)
def _(stl, depth=0):
depth += 1
t = sym.Symbol("t{}".format(depth))
rl, rh = smooth_robustness(stl.arg)
r = (rl.subs({t_sym: t}), rh.subs({t_sym: t}))
return -F2(r[1], stl.interval, t), -F1(r[0], stl.interval, t)
los, his = zip(*[rl.subs({t_sym: t}), rh.subs({t_sym: t}) for t in times])
return LSE(rl), soft_max(map(x_ij(L, H), his))
@smooth_robustness.register(stl.Neg)
def _(stl, depth=0):
def _(stl, L, H):
rl, rh = smooth_robustness(arg)
return -rh, -rl
@ -71,7 +63,7 @@ op_lookup = {
@smooth_robustness.register(stl.LinEq)
def _(stl, depth=0):
def _(stl, L, H):
op = op_lookup[stl.op]
retval = op(eval_terms(stl), stl.const)
return retval, retval