def find_hl(data, domain, active_indices, solver):
# Constants
n_r = len(domain.real_vars)
real_features = [[row[v] for v in domain.real_vars] for row, _ in data]
labels = [row[1] for row in data]
# Variables
a_r = [smt.Symbol("a_r[{}]".format(r), REAL) for r in range(n_r)]
b = smt.Symbol("b", REAL)
# Constraints
for i in active_indices:
x_r, label = real_features[i], labels[i]
sum_coefficients = smt.Plus(
[a_r[r] * smt.Real(x_r[r]) for r in range(n_r)])
if label:
solver.add_assertion(sum_coefficients + DELTA <= b)
else:
solver.add_assertion(sum_coefficients - DELTA > b)
if not solver.solve():
return None
model = solver.get_model()
x_vars = [domain.get_symbol(domain.real_vars[r]) for r in range(n_r)]
return smt.Plus([model.get_value(a_r[r]) * x_vars[r]
for r in range(n_r)]) <= model.get_value(b)
def generate_half_space_sample(domain, real_count):
samples = [get_sample(domain) for _ in range(real_count)]
coefficients, offset = Learner.fit_hyperplane(domain, samples)
coefficients = [smt.Real(float(coefficients[i][0])) * domain.get_symbol(domain.real_vars[i]) for i in
range(real_count)]
if random.random() < 0.5:
return smt.Plus(*coefficients) <= offset
else:
return smt.Plus(*coefficients) >= offset
def ast_to_smt(self, node):
"""
:type node: Node
"""
def convert_children(number=None):
if number is not None and len(node.children) != number:
raise Exception(
"The number of children ({}) differed from {}".format(
len(node.children), number))
return [self.ast_to_smt(child) for child in node.children]
if node.name == "ite":
return smt.Ite(*convert_children(3))
elif node.name == "~":
return smt.Not(*convert_children(1))
elif node.name == "^":
return smt.Pow(*convert_children(2))
elif node.name == "&":
return smt.And(*convert_children())
elif node.name == "|":
return smt.Or(*convert_children())
elif node.name == "*":
return smt.Times(*convert_children())
elif node.name == "+":
return smt.Plus(*convert_children())
elif node.name == "-":
return smt.Minus(*convert_children(2))
elif node.name == "<=":
return smt.LE(*convert_children(2))
elif node.name == ">=":
return smt.GE(*convert_children(2))
elif node.name == "<":
return smt.LT(*convert_children(2))
elif node.name == ">":
return smt.GT(*convert_children(2))
elif node.name == "=":
return smt.Equals(*convert_children(2))
elif node.name == "const":
c_type, c_value = [child.name for child in node.children]
if c_type == "bool":
return smt.Bool(bool(c_value))
elif c_type == "real":
return smt.Real(float(c_value))
else:
raise Exception("Unknown constant type {}".format(c_type))
elif node.name == "var":
v_type, v_name = [child.name for child in node.children]
if v_type == "bool":
v_smt_type = smt.BOOL
elif v_type == "real":
v_smt_type = smt.REAL
else:
raise Exception("Unknown variable type {}".format(v_type))
return smt.Symbol(v_name, v_smt_type)
else:
raise RuntimeError("Unrecognized node type '{}'".format(node.name))
def _exp_to_smt(self, expression):
if isinstance(expression, sympy.Add):
return smt.Plus([self._exp_to_smt(arg) for arg in expression.args])
elif isinstance(expression, sympy.Mul):
return smt.Times(*[self._exp_to_smt(arg) for arg in expression.args])
elif isinstance(expression, sympy.Symbol):
return smt.Symbol(str(expression), INT)
try:
expression = int(expression)
return smt.Int(expression)
except ValueError:
pass
raise RuntimeError("Could not parse {} of type {}".format(expression, type(expression)))
def dual_paths_distinct(n):
booleans = [] # ["A{}".format(i) for i in range(n)]
domain = Domain.make(booleans, ["x{}".format(i) for i in range(n)], real_bounds=(0, 1))
bool_vars = domain.get_bool_symbols()
real_vars = domain.get_real_symbols()
terms = []
for i in range(n):
v1, v2 = random.sample(real_vars, 2)
terms.append(v1 * random.random() <= v2 * random.random())
paths = []
for i in range(n):
paths.append(smt.Ite(smt.And(*random.sample(bool_vars + terms, n)), smt.Real(i), smt.Real(0)))
return Density(domain, domain.get_bounds(), smt.Plus(*paths))
def _test_to_smt(self, operator):
operator = operator.to_canonical()
# FIXME Integer rounding only applicable if x >= 0
def to_symbol(s):
return smt.Symbol(s, typename=smt.types.INT)
import math
items = [smt.Times(smt.Int(int(math.floor(v))), to_symbol(k)) for k, v in operator.lhs.items()]
lhs = smt.Plus([smt.Int(0)] + items)
rhs = smt.Int(int(math.floor(operator.rhs)))
assert operator.symbol == "<="
return smt.LE(lhs, rhs)
def affine(self, a):
""" return some positive rescaling of the affine expression
s.t. the rescaled expression has integer coefficients
safe, since positive rescaling preserves
all of a >= 0, a > 0, and a == 0 """
# find the lcm of the offset denominator
# and all coefficient denominators
mult = a.offset.denominator
for t in a.terms:
mult = _lcm(mult, t.coeff.denominator)
# now, we can produce an integral affine equation,
# by rescaling through with `mult`
a = a * Fraction(mult)
# Finally, convert this to an SMT formula
assert a.offset.denominator == 1
f = SMT.Int(a.offset.numerator)
for t in a.terms:
assert t.coeff.denominator == 1
sym = self._ctxt.get(t.var)
assert sym is not None, f"expected variable name '{t.var}'"
term = SMT.Times(SMT.Int(t.coeff.numerator), sym)
f = SMT.Plus(f, term)
return f
def walk_plus(self, args):
return smt.Plus(*self.walk_smt_multiple(args))
def balancer_flow_formula(junctions, width, length):
formula = []
belts = [[
Belt(s.Symbol(f'b{i}[{x}].rho', t.REAL),
s.Symbol(f'b{i}[{x}].v', t.REAL)) for x in range(length + 1)
] for i in range(width)]
for beltway in belts:
for belt in beltway:
formula.extend(domain(belt))
if not s.is_sat(s.And(formula)):
raise Exception('Domain is not SAT :/')
# Balancing rules.
junctions_by_x = [[] for x in range(length + 1)]
for (x, y1, y2) in junctions:
junctions_by_x[x].append((y1, y2))
inn = x - 1
out = x
input_rho = s.Plus(belts[y1][inn].rho, belts[y2][inn].rho)
# We want to put half of the input on each output.
half_input = s.Div(input_rho, s.Real(2))
# If output velocity is less than the half_input that we would like to
# assign to it, we've filled it. Velocity is a hard limit, because it's out
# of influence of this splitter. We'll set the output density to 1 in that
# case. Aside: The flux is the min of that and the velocity, so if
# out-velocity is the limiting factor, it won't change the flux calculation
# to just assign rho_out = v_out.
#
# Now, the excess that we couldn't assign has to go somewhere: (1) to the
# other output belt; if that's full, (2) feed back up the chain by reducing
# input velocities.
excess_from_1 = s.Max(s.Real(0), s.Minus(half_input, belts[y1][out].v))
excess_from_2 = s.Max(s.Real(0), s.Minus(half_input, belts[y2][out].v))
# This formula is most accurate for assignment > velocity (density will
# equal 1), but it doesn't change the flux calculation just toset rho to
# the velocity when velocity limits flow. (So you should be able to replace
# the Ite by v_out and be OK.)
formula.append(
s.Equals(
belts[y1][out].rho,
s.Ite(half_input + excess_from_2 > belts[y1][out].v, s.Real(1),
half_input + excess_from_2)))
formula.append(
s.Equals(
belts[y2][out].rho,
s.Ite(half_input + excess_from_1 > belts[y2][out].v, s.Real(1),
half_input + excess_from_1)))
output_v = s.Plus(belts[y1][out].v, belts[y2][out].v)
half_output = s.Div(output_v, s.Real(2))
unused_density_from_1 = s.Max(s.Real(0),
s.Minus(half_output, belts[y1][inn].rho))
unused_density_from_2 = s.Max(s.Real(0),
s.Minus(half_output, belts[y2][inn].rho))
formula.append(
s.Equals(
belts[y1][inn].v,
s.Ite(half_output + unused_density_from_2 > belts[y1][inn].rho,
s.Real(1), half_output + unused_density_from_2)))
formula.append(
s.Equals(
belts[y2][inn].v,
s.Ite(half_output + unused_density_from_1 > belts[y2][inn].rho,
s.Real(1), half_output + unused_density_from_1)))
# Conservation of flux at each junction.
input_flux = s.Plus(belts[y1][inn].flux, belts[y2][inn].flux)
output_flux = s.Plus(belts[y1][out].flux, belts[y2][out].flux)
formula.append(s.Equals(input_flux, output_flux))
# Any belts not involved in a junction are pass-throughs. Their successive
# values must remain equal.
thru_belts = [
list(
set(range(width)) - {y1
for y1, y2 in junctions_by_x[x]} -
{y2
for y1, y2 in junctions_by_x[x]}) for x in range(length + 1)
]
for x, thru in enumerate(thru_belts[1:]):
for y in thru:
formula.append(s.Equals(belts[y][x].rho, belts[y][x + 1].rho))
formula.append(s.Equals(belts[y][x].v, belts[y][x + 1].v))
return formula, belts
def prove_properties(junctions):
width = max(max(y1 for _, y1, _ in junctions),
max(y2 for _, _, y2 in junctions)) + 1
length = max(x for x, _, _ in junctions)
formula, belts = balancer_flow_formula(junctions, width, length)
supply = s.Plus(beltway[0].rho for beltway in belts)
demand = s.Plus(beltway[-1].v for beltway in belts)
max_theoretical_throughput = s.Min(supply, demand)
actual_throughput = s.Plus(beltway[-1].flux for beltway in belts)
fully_balanced_output = []
max_flux = s.Max(b[-1].flux for b in belts)
for i, b1 in enumerate(belts):
fully_balanced_output.append(
s.Or(s.Equals(b1[-1].flux, max_flux),
s.Equals(b1[-1].flux, b1[-1].v)))
fully_balanced_output = s.And(fully_balanced_output)
fully_balanced_input = []
max_flux = s.Max(b[0].flux for b in belts)
for i, b1 in enumerate(belts):
fully_balanced_input.append(
s.Or(s.Equals(b1[0].flux, max_flux),
s.Equals(b1[0].flux, b1[0].rho)))
fully_balanced_input = s.And(fully_balanced_input)
with s.Solver() as solver:
solver.add_assertion(s.And(formula))
solver.push()
solver.add_assertion(
s.Not(s.Equals(max_theoretical_throughput, actual_throughput)))
if not solver.solve():
print("It's throughput-unlimited!")
else:
print("It's throughput-limited; here's an example:")
m = solver.get_model()
inputs = tuple(m.get_py_value(beltway[0].rho) for beltway in belts)
outputs = tuple(m.get_py_value(beltway[-1].v) for beltway in belts)
print(f'Input lane densities: ({", ".join(map(str, inputs))})')
print(f'Output lane velocities: ({", ".join(map(str, outputs))})')
check_balancer_flow(junctions, inputs, outputs)
solver.pop()
solver.push()
solver.add_assertion(s.Not(fully_balanced_input))
if not solver.solve():
print("It's input-balanced!")
else:
print("It's input-imbalanced; here's an example:")
m = solver.get_model()
inputs = tuple(m.get_py_value(beltway[0].rho) for beltway in belts)
outputs = tuple(m.get_py_value(beltway[-1].v) for beltway in belts)
print(f'Input lane densities: ({", ".join(map(str, inputs))})')
print(f'Output lane velocities: ({", ".join(map(str, outputs))})')
check_balancer_flow(junctions, inputs, outputs)
solver.pop()
solver.push()
solver.add_assertion(s.Not(fully_balanced_output))
if solver.solve:
print("It's output-balanced!")
else:
print("It's output-imbalanced; here's an example:")
m = solver.get_model()
inputs = tuple(m.get_py_value(beltway[0].rho) for beltway in belts)
outputs = tuple(m.get_py_value(beltway[-1].v) for beltway in belts)
print(f'Input lane densities: ({", ".join(map(str, inputs))})')
print(f'Output lane velocities: ({", ".join(map(str, outputs))})')
check_balancer_flow(junctions, inputs, outputs)
def generate_half_space(domain, real_count):
coefficients = [smt.Real(random.random() * 2 - 1) * domain.get_symbol(domain.real_vars[i]) for i in
range(real_count)]
return smt.LE(smt.Plus(*coefficients), smt.Real(random.random() * 2 - 1))
def __getExpressionTree(symbolicExpression):
# TODO LATER: take into account bitwise shift operations
args = []
castType = None
if len(symbolicExpression.args) > 0:
for symbolicArg in symbolicExpression.args:
arg, type = Solver.__getExpressionTree(symbolicArg)
args.append(arg)
if castType is None:
castType = type
else:
if castType.literal == 'Integer':
if type.literal == 'Real':
castType = type
# TODO LATER: consider other possible castings
if castType.literal == 'Real':
for i in range(len(args)):
args[i] = pysmt.ToReal(args[i])
if isinstance(symbolicExpression, sympy.Not):
if castType.literal == 'Integer':
return pysmt.Equals(args[0], pysmt.Int(0)), Type('Bool')
elif castType.literal == 'Real':
return pysmt.Equals(args[0], pysmt.Real(0)), Type('Bool')
elif castType.literal == 'Bool':
return pysmt.Not(args[0]), Type('Bool')
else: # castType.literal == 'BitVector'
return pysmt.BVNot(args[0]), Type('BitVector')
elif isinstance(symbolicExpression, sympy.Lt):
return pysmt.LT(args[0], args[1]), Type('Bool')
elif isinstance(symbolicExpression, sympy.Gt):
return pysmt.GT(args[0], args[1]), Type('Bool')
elif isinstance(symbolicExpression, sympy.Ge):
return pysmt.GE(args[0], args[1]), Type('Bool')
elif isinstance(symbolicExpression, sympy.Le):
return pysmt.LE(args[0], args[1]), Type('Bool')
elif isinstance(symbolicExpression, sympy.Eq):
return pysmt.Equals(args[0], args[1]), Type('Bool')
elif isinstance(symbolicExpression, sympy.Ne):
return pysmt.NotEquals(args[0], args[1]), Type('Bool')
elif isinstance(symbolicExpression, sympy.And):
if castType.literal == 'Bool':
return pysmt.And(args[0], args[1]), Type('Bool')
else: # type.literal == 'BitVector'
return pysmt.BVAnd(args[0], args[1]), castType
elif isinstance(symbolicExpression, sympy.Or):
if castType.literal == 'Bool':
return pysmt.Or(args[0], args[1]), Type('Bool')
else: # type.literal == 'BitVector'
return pysmt.BVOr(args[0], args[1]), castType
elif isinstance(symbolicExpression, sympy.Xor):
return pysmt.BVXor(args[0], args[1]), castType
elif isinstance(symbolicExpression, sympy.Add):
return pysmt.Plus(args), castType
elif isinstance(symbolicExpression, sympy.Mul):
return pysmt.Times(args), castType
elif isinstance(symbolicExpression, sympy.Pow):
return pysmt.Pow(args[0], args[1]), castType
# TODO LATER: deal with missing modulo operator from pysmt
else:
if isinstance(symbolicExpression, sympy.Symbol):
symbolType = Variable.symbolTypes[symbolicExpression.name]
literal = symbolType.getTypeForSolver()
designator = symbolType.designatorExpr1
type = Type(literal, designator)
return Solver.__encodeTerminal(symbolicExpression, type), type
elif isinstance(symbolicExpression, sympy.Integer):
type = Type('Integer')
return Solver.__encodeTerminal(symbolicExpression, type), type
elif isinstance(symbolicExpression, sympy.Rational):
type = Type('Real')
return Solver.__encodeTerminal(symbolicExpression, type), type
elif isinstance(symbolicExpression, sympy.Float):
type = Type('Real')
return Solver.__encodeTerminal(symbolicExpression, type), type
else:
type = Type('Real')
return Solver.__encodeTerminal(symbolicExpression, type), type
def learn_partial(self, solver, domain, data, new_active_indices):
# Constants
n_b_original = len(domain.bool_vars)
n_b = n_b_original * 2
n_r = len(domain.real_vars)
n_h_original = self.half_space_count if n_r > 0 else 0
n_h = n_h_original * 2 if self.allow_negations else n_h_original
n_c = self.conjunction_count
n_d = len(data)
real_features = [[row[v] for v in domain.real_vars] for row, _ in data]
bool_features = [[row[v] for v in domain.bool_vars] for row, _ in data]
labels = [row[1] for row in data]
# Variables
a_hr = [[
smt.Symbol("a_hr[{}][{}]".format(h, r), REAL) for r in range(n_r)
] for h in range(n_h_original)]
b_h = [
smt.Symbol("b_h[{}]".format(h), REAL) for h in range(n_h_original)
]
s_ch = [[smt.Symbol("s_ch[{}][{}]".format(c, h)) for h in range(n_h)]
for c in range(n_c)]
s_cb = [[smt.Symbol("s_cb[{}][{}]".format(c, b)) for b in range(n_b)]
for c in range(n_c)]
# Aux variables
s_ih = [[smt.Symbol("s_ih[{}][{}]".format(i, h)) for h in range(n_h)]
for i in range(n_d)]
s_ic = [[smt.Symbol("s_ic[{}][{}]".format(i, c)) for c in range(n_c)]
for i in range(n_d)]
# Constraints
for i in new_active_indices:
x_r, x_b, label = real_features[i], bool_features[i], labels[i]
for h in range(n_h_original):
sum_coefficients = smt.Plus(
[a_hr[h][r] * smt.Real(x_r[r]) for r in range(n_r)])
solver.add_assertion(
smt.Iff(s_ih[i][h], sum_coefficients <= b_h[h]))
for h in range(n_h_original, n_h):
solver.add_assertion(
smt.Iff(s_ih[i][h], ~s_ih[i][h - n_h_original]))
for c in range(n_c):
solver.add_assertion(
smt.Iff(
s_ic[i][c],
smt.And([smt.TRUE()] + [(~s_ch[c][h] | s_ih[i][h])
for h in range(n_h)] +
[
~s_cb[c][b]
for b in range(n_b_original) if not x_b[b]
] + [
~s_cb[c][b]
for b in range(n_b_original, n_b)
if x_b[b - n_b_original]
])))
if label:
solver.add_assertion(smt.Or([s_ic[i][c] for c in range(n_c)]))
else:
solver.add_assertion(smt.And([~s_ic[i][c]
for c in range(n_c)]))
solver.solve()
model = solver.get_model()
x_vars = [domain.get_symbol(domain.real_vars[r]) for r in range(n_r)]
half_spaces = [
smt.Plus(
[model.get_value(a_hr[h][r]) * x_vars[r]
for r in range(n_r)]) <= model.get_value(b_h[h])
for h in range(n_h_original)
] + [
smt.Plus(
[model.get_value(a_hr[h][r]) * x_vars[r]
for r in range(n_r)]) > model.get_value(b_h[h])
for h in range(n_h - n_h_original)
]
b_vars = [
domain.get_symbol(domain.bool_vars[b]) for b in range(n_b_original)
]
bool_literals = [b_vars[b] for b in range(n_b_original)]
bool_literals += [~b_vars[b] for b in range(n_b - n_b_original)]
conjunctions = [[
half_spaces[h]
for h in range(n_h) if model.get_py_value(s_ch[c][h])
] + [
bool_literals[b]
for b in range(n_b) if model.get_py_value(s_cb[c][b])
] for c in range(n_c)]
return smt.Or([smt.And(conjunction) for conjunction in conjunctions])
SMT.GT(variables[i + 2], SMT.Int(0))),
SMT.Equals(variables[i + 1], SMT.Int(1))))
solver = SMT.Solver(name="z3")
print("constraints:")
for c in constraints:
print(c)
solver.add_assertion(c)
# add equations
# 24727*a_1 + 75235*b_1 + 50508*c_1 = 75235*a_2 + 125743*b_2 + 176251*c_2
solver.add_assertion(
SMT.Equals(
SMT.Plus(
SMT.Plus(SMT.Times(SMT.Int(125743), variables[4]),
SMT.Times(SMT.Int(75235), variables[3])),
SMT.Times(SMT.Int(176251), variables[5])),
SMT.Plus(
SMT.Plus(SMT.Times(SMT.Int(75235), variables[1]),
SMT.Times(SMT.Int(24727), variables[0])),
SMT.Times(SMT.Int(50508), variables[2]))))
# 75235*a_2 + 125743*b_2 + 176251*c_2 = 125743*a_3 + 301994*b_3 + 16785921*c
solver.add_assertion(
SMT.Equals(
SMT.Plus(
SMT.Plus(SMT.Times(SMT.Int(301994), variables[7]),
SMT.Times(SMT.Int(125743), variables[6])),
SMT.Times(SMT.Int(16785921), variables[8])),
SMT.Plus(
SMT.Plus(SMT.Times(SMT.Int(125743), variables[4]),
def learn_partial(self, solver, domain: Domain, data: np.ndarray,
labels: np.ndarray, new_active_indices: Set):
# Constants
n_b_original = len(domain.bool_vars)
n_b = n_b_original * 2
n_r = len(domain.real_vars)
n_h_original = self.half_space_count if n_r > 0 else 0
n_h = n_h_original * 2 if self.allow_negations else n_h_original
n_c = self.conjunction_count
n_d = data.shape[0]
real_indices = np.array(
[domain.var_types[v] == smt.REAL for v in domain.variables])
real_features = data[:, real_indices]
bool_features = data[:, np.logical_not(real_indices)]
# Variables
a_hr = [[
smt.Symbol("a_hr[{}][{}]".format(h, r), REAL) for r in range(n_r)
] for h in range(n_h_original)]
b_h = [
smt.Symbol("b_h[{}]".format(h), REAL) for h in range(n_h_original)
]
s_ch = [[smt.Symbol("s_ch[{}][{}]".format(c, h)) for h in range(n_h)]
for c in range(n_c)]
s_cb = [[smt.Symbol("s_cb[{}][{}]".format(c, b)) for b in range(n_b)]
for c in range(n_c)]
# Aux variables
s_ih = [[smt.Symbol("s_ih[{}][{}]".format(i, h)) for h in range(n_h)]
for i in range(n_d)]
s_ic = [[smt.Symbol("s_ic[{}][{}]".format(i, c)) for c in range(n_c)]
for i in range(n_d)]
def pair(real: bool, c: int, index: int) -> Tuple[FNode, FNode]:
if real:
return s_ch[c][index], s_ch[c][index + n_h_original]
else:
return s_cb[c][index], s_cb[c][index + n_b_original]
def order_equal(pair1, pair2):
x_t, x_f, y_t, y_f = pair1 + pair2
return smt.Iff(x_t, y_t) & smt.Iff(x_f, y_f)
def order_geq(pair1, pair2):
x_t, x_f, y_t, y_f = pair1 + pair2
return x_t | y_f | ((~x_f) & (~y_t))
def pairs(c: int) -> List[Tuple[FNode, FNode]]:
return [pair(True, c, i) for i in range(n_h_original)
] + [pair(False, c, i) for i in range(n_b_original)]
def order_geq_lex(c1: int, c2: int):
pairs_c1, pairs_c2 = pairs(c1), pairs(c2)
assert len(pairs_c1) == len(pairs_c2)
constraints = smt.TRUE()
for j in range(len(pairs_c1)):
condition = smt.TRUE()
for i in range(j):
condition &= order_equal(pairs_c1[i], pairs_c2[i])
constraints &= smt.Implies(condition,
order_geq(pairs_c1[j], pairs_c2[j]))
return constraints
# Constraints
for i in new_active_indices:
x_r, x_b, label = [float(val) for val in real_features[i]
], bool_features[i], labels[i]
for h in range(n_h_original):
sum_coefficients = smt.Plus(
[a_hr[h][r] * smt.Real(x_r[r]) for r in range(n_r)])
solver.add_assertion(
smt.Iff(s_ih[i][h], sum_coefficients <= b_h[h]))
for h in range(n_h_original, n_h):
solver.add_assertion(
smt.Iff(s_ih[i][h], ~s_ih[i][h - n_h_original]))
for c in range(n_c):
solver.add_assertion(
smt.Iff(
s_ic[i][c],
smt.
Or([smt.FALSE()] + [(s_ch[c][h] & s_ih[i][h])
for h in range(n_h)] +
[s_cb[c][b]
for b in range(n_b_original) if x_b[b]] + [
s_cb[c][b] for b in range(n_b_original, n_b)
if not x_b[b - n_b_original]
])))
# --- [start] symmetry breaking ---
# Mutually exclusive
if "m" in self.symmetries:
for c in range(n_c):
for h in range(n_h_original):
solver.add_assertion(~(s_ch[c][h]
& s_ch[c][h + n_h_original]))
for b in range(n_b_original):
solver.add_assertion(~(s_cb[c][b]
& s_cb[c][b + n_b_original]))
# Normalized
if "n" in self.symmetries:
for h in range(n_h_original):
solver.add_assertion(
smt.Equals(b_h[h], smt.Real(1.0))
| smt.Equals(b_h[h], smt.Real(0.0)))
# Vertical symmetries
if "v" in self.symmetries:
for c in range(n_c - 1):
solver.add_assertion(order_geq_lex(c, c + 1))
# Horizontal symmetries
if "h" in self.symmetries:
for h in range(n_h_original - 1):
solver.add_assertion(a_hr[h][0] >= a_hr[h + 1][0])
# --- [end] symmetry breaking ---
if label:
solver.add_assertion(smt.And([s_ic[i][c] for c in range(n_c)]))
else:
solver.add_assertion(smt.Or([~s_ic[i][c] for c in range(n_c)]))
solver.solve()
model = solver.get_model()
x_vars = [domain.get_symbol(domain.real_vars[r]) for r in range(n_r)]
half_spaces = [
smt.Plus(
[model.get_value(a_hr[h][r]) * x_vars[r]
for r in range(n_r)]) <= model.get_value(b_h[h])
for h in range(n_h_original)
] + [
smt.Plus(
[model.get_value(a_hr[h][r]) * x_vars[r]
for r in range(n_r)]) > model.get_value(b_h[h])
for h in range(n_h - n_h_original)
]
b_vars = [
domain.get_symbol(domain.bool_vars[b]) for b in range(n_b_original)
]
bool_literals = [b_vars[b] for b in range(n_b_original)]
bool_literals += [~b_vars[b] for b in range(n_b - n_b_original)]
conjunctions = [[
half_spaces[h]
for h in range(n_h) if model.get_py_value(s_ch[c][h])
] + [
bool_literals[b]
for b in range(n_b) if model.get_py_value(s_cb[c][b])
] for c in range(n_c)]
return smt.And([smt.Or(conjunction) for conjunction in conjunctions])
def find_cnf(data, domain, active_indices, solver, n_c, n_h):
# Constants
n_b_original = len(domain.bool_vars)
n_b = n_b_original * 2
n_r = len(domain.real_vars)
n_d = len(data)
real_features = [[row[v] for v in domain.real_vars] for row, _ in data]
bool_features = [[row[v] for v in domain.bool_vars] for row, _ in data]
labels = [row[1] for row in data]
# Variables
a_hr = [[
smt.Symbol("a_hr[{}][{}]".format(h, r), REAL) for r in range(n_r)
] for h in range(n_h)]
b_h = [smt.Symbol("b_h[{}]".format(h), REAL) for h in range(n_h)]
s_ch = [[smt.Symbol("s_ch[{}][{}]".format(c, h)) for h in range(n_h)]
for c in range(n_c)]
s_cb = [[smt.Symbol("s_cb[{}][{}]".format(c, b)) for b in range(n_b)]
for c in range(n_c)]
# Aux variables
s_ih = [[smt.Symbol("s_ih[{}][{}]".format(i, h)) for h in range(n_h)]
for i in range(n_d)]
s_ic = [[smt.Symbol("s_ic[{}][{}]".format(i, c)) for c in range(n_c)]
for i in range(n_d)]
# Constraints
for i in active_indices:
x_r, x_b, label = real_features[i], bool_features[i], labels[i]
for h in range(n_h):
sum_coefficients = smt.Plus(
[a_hr[h][r] * smt.Real(x_r[r]) for r in range(n_r)])
if label:
solver.add_assertion(
smt.Iff(s_ih[i][h], sum_coefficients + DELTA <= b_h[h]))
else:
solver.add_assertion(
smt.Iff(s_ih[i][h], sum_coefficients - DELTA <= b_h[h]))
for c in range(n_c):
solver.add_assertion(
smt.Iff(
s_ic[i][c],
smt.Or([smt.FALSE()] + [(s_ch[c][h] & s_ih[i][h])
for h in range(n_h)] +
[s_cb[c][b]
for b in range(n_b_original) if x_b[b]] + [
s_cb[c][b] for b in range(n_b_original, n_b)
if not x_b[b - n_b_original]
])))
if label:
solver.add_assertion(smt.And([s_ic[i][c] for c in range(n_c)]))
else:
solver.add_assertion(smt.Or([~s_ic[i][c] for c in range(n_c)]))
if not solver.solve():
return None
model = solver.get_model()
x_vars = [domain.get_symbol(domain.real_vars[r]) for r in range(n_r)]
half_spaces = [
smt.Plus([model.get_value(a_hr[h][r]) * x_vars[r]
for r in range(n_r)]) <= model.get_value(b_h[h])
for h in range(n_h)
]
b_vars = [
domain.get_symbol(domain.bool_vars[b]) for b in range(n_b_original)
]
bool_literals = [b_vars[b] for b in range(n_b_original)]
bool_literals += [~b_vars[b] for b in range(n_b - n_b_original)]
conjunctions = [
[half_spaces[h]
for h in range(n_h) if model.get_py_value(s_ch[c][h])] + [
bool_literals[b]
for b in range(n_b) if model.get_py_value(s_cb[c][b])
] for c in range(n_c)
]
return smt.And([smt.Or(conjunction) for conjunction in conjunctions])
def initialize(self, mdp, colors, hole_options, reward_name, okay_states,
target_states, threshold, relation):
logger.warning("This approach has been tested sparsely.")
prob0E, prob1A = stormpy.compute_prob01min_states(
mdp, okay_states, target_states)
sink_states = ~okay_states
assert len(mdp.initial_states) == 1
self.state_vars = [
smt.Symbol("p_{}".format(i), smt.REAL)
for i in range(mdp.nr_states)
]
self.state_prob1_vars = [
smt.Symbol("asure_{}".format(i)) for i in range(mdp.nr_states)
]
self.state_probpos_vars = [
smt.Symbol("x_{}".format(i)) for i in range(mdp.nr_states)
]
self.state_order_vars = [
smt.Symbol("r_{}".format(i), smt.REAL)
for i in range(mdp.nr_states)
]
self.option_vars = dict()
for h, opts in hole_options.items():
self.option_vars[h] = {
index: smt.Symbol("h_{}_{}".format(h, opt))
for index, opt in enumerate(opts)
}
self.transition_system = []
logger.debug("Obtain rewards if necessary")
rewards = mdp.reward_models[reward_name] if reward_name else None
if rewards:
assert not rewards.has_transition_rewards
state_rewards = rewards.has_state_rewards
action_rewards = rewards.has_state_action_rewards
logger.debug(
"Model has state rewards: {}, has state/action rewards {}".
format(state_rewards, action_rewards))
self.transition_system.append(
self.state_prob1_vars[mdp.initial_states[0]])
threshold_inequality, action_constraint_inequality = self._to_smt_relation(
relation) # TODO or GE
self.transition_system.append(
threshold_inequality(self.state_vars[mdp.initial_states[0]],
smt.Real(float(threshold))))
state_order_domain_constraint = smt.And([
smt.And(smt.GE(var, smt.Real(0)), smt.LE(var, smt.Real(1)))
for var in self.state_order_vars
])
self.transition_system.append(state_order_domain_constraint)
#TODO how to ensure that prob is zero if there is no path.
select_parameter_value_constraints = []
for h, opts in self.option_vars.items():
select_parameter_value_constraints.append(
smt.Or(ov for ov in opts.values()))
for i, opt1 in enumerate(opts.values()):
for opt2 in list(opts.values())[i + 1:]:
select_parameter_value_constraints.append(
smt.Not(smt.And(opt1, opt2)))
#logger.debug("Consistency: {}".format(smt.And(select_parameter_value_constraints)))
self.transition_system.append(
smt.And(select_parameter_value_constraints))
for state in mdp.states:
if sink_states.get(state.id):
assert rewards is None
self.transition_system.append(
smt.Equals(self.state_vars[state.id], smt.REAL(0)))
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
self.transition_system.append(
smt.Not(self.state_prob1_vars[state.id]))
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
self.transition_system.append(
smt.Equals(self.state_order_vars[state.id], smt.Real(0)))
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
elif target_states.get(state.id):
self.transition_system.append(
smt.Equals(self.state_order_vars[state.id], smt.Real(1)))
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
self.transition_system.append(self.state_prob1_vars[state.id])
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
if rewards is None:
self.transition_system.append(
smt.Equals(self.state_vars[state.id], smt.Real(1)))
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
else:
self.transition_system.append(
self.state_probpos_vars[state.id])
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
self.transition_system.append(
smt.Equals(self.state_vars[state.id], smt.Real(0)))
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
else:
state_in_prob1A = False
state_in_prob0E = False
if prob0E.get(state.id):
state_in_prob0E = True
else:
self.transition_system.append(
smt.Equals(self.state_order_vars[state.id],
smt.Real(1)))
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
self.transition_system.append(
self.state_probpos_vars[state.id])
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
if rewards and not state_in_prob0E:
if prob1A.get(state.id):
self.transition_system.append(
self.state_prob1_vars[state.id])
logger.debug("Constraint: {}".format(
self.transition_system[-1]))
state_in_prob1A = True
for action in state.actions:
action_index = mdp.nondeterministic_choice_indices[
state.id] + action.id
#logger.debug("Action index: {}".format(action_index))
precondition = smt.And([
self.option_vars[hole][list(option)[0]] for hole,
option in colors.get(action_index, dict()).items()
])
reward_value = None
if rewards:
reward_const = (rewards.get_state_reward(
state.id) if state_rewards else 0.0) + (
rewards.get_state_action_reward(action_index)
if action_rewards else 0.0)
reward_value = smt.Real(reward_const)
act_constraint = action_constraint_inequality(
self.state_vars[state.id],
smt.Plus([
smt.Times(smt.Real(t.value()), self.
state_vars[t.column])
for t in action.transitions
] + [reward_value] if reward_value else []))
full_act_constraint = act_constraint
if state_in_prob0E:
if not rewards:
full_act_constraint = smt.And(
smt.Implies(self.state_probpos_vars[state.id],
act_constraint),
smt.Implies(
smt.Not(self.state_probpos_vars),
smt.Equals(self.state_vars[state.id],
smt.Real(0))))
self.transition_system.append(
smt.Implies(
precondition,
smt.Iff(
self.state_probpos_vars[state.id],
smt.Or([
smt.And(
self.state_probpos_vars[t.column],
smt.LT(
self.state_order_vars[
state.id],
self.state_order_vars[
t.column]))
for t in action.transitions
]))))
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
if rewards and not state_in_prob1A:
# prob_one(state) <-> probpos AND for all succ prob_one(succ)
self.transition_system.append(
smt.Implies(
precondition,
smt.Iff(
self.state_prob1_vars[state.id],
smt.And([
self.state_prob1_vars[t.column]
for t in action.transitions
] + [self.state_probpos_vars[state.id]]))))
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
self.transition_system.append(
smt.Implies(precondition, full_act_constraint))
#logger.debug("Constraint: {}".format(self.transition_system[-1]))
if rewards:
self.transition_system.append(
smt.And([smt.GE(sv, smt.Real(0)) for sv in self.state_vars]))
else:
self.transition_system.append(
smt.And([
smt.And(smt.GE(sv, smt.Real(0)), smt.LE(sv, smt.Real(1)))
for sv in self.state_vars
]))
#print(self.transition_system)
formula = smt.And(self.transition_system)
logger.info("Start SMT solver")
model = smt.get_model(formula)
if model:
logger.info("SAT: Found {}".format(model))
else:
logger.info("UNSAT.")
