def _generate_episode_reinforce(self, sample, max_flips, walk_prob):
f = sample.formula
data = init_tensors(sample, self.device)
true_lit_count = compute_true_lit_count(f.clauses, data.sol)
log_probs = []
flip = 0
flipped = set()
backflipped = 0
unsat_clauses = []
while flip < max_flips:
unsat_clause_indices = [
k for k in range(len(f.clauses)) if true_lit_count[k] == 0
]
unsat_clauses.append(len(unsat_clause_indices))
# print(len(unsat_clause_indices))
sat = not unsat_clause_indices
if sat:
break
if random.random() < walk_prob:
unsat_clause = f.clauses[random.choice(unsat_clause_indices)]
v, log_prob = abs(random.choice(unsat_clause)) - 1, None
else:
v, log_prob = self._select_variable_reinforce(data)
if v.item() not in flipped:
flipped.add(v.item())
else:
backflipped += 1
flip_(data.x[0], data.sol, true_lit_count, v, f.occur_list)
flip += 1
log_probs.append(log_prob)
# print(backflipped)
return sat, (flip, backflipped, unsat_clauses), (log_probs, )
def _generate_episode_a2c(self, sample, max_flips):
f = sample.formula
data = init_tensors(sample, self.device)
true_lit_count = compute_true_lit_count(f.clauses, data.sol)
log_probs = []
values = []
entropies = []
flip = 0
while flip < max_flips:
unsat_clause_indices = [
k for k in range(len(f.clauses)) if true_lit_count[k] == 0
]
sat = not unsat_clause_indices
if sat:
break
v, log_prob, value, entropy = self._select_variable_a2c(data)
flip_(data.x[0], data.sol, true_lit_count, v, f.occur_list)
flip += 1
log_probs.append(log_prob)
values.append(value)
entropies.append(entropy)
return sat, flip, (log_probs, values, entropies)
def _eval_generate_episode_reinforce(self, sample, max_flips, walk_prob):
f = sample.formula
data = init_tensors(sample, self.device)
true_lit_count = compute_true_lit_count(f.clauses, data.sol)
flip = 0
while flip < max_flips:
unsat_clause_indices = [
k for k in range(len(f.clauses)) if true_lit_count[k] == 0
]
sat = not unsat_clause_indices
if sat:
break
if random.random() < walk_prob:
unsat_clause = f.clauses[random.choice(unsat_clause_indices)]
v = abs(random.choice(unsat_clause)) - 1
else:
v = self._eval_select_variable_reinforce(data)
flip_(data.x[0], data.sol, true_lit_count, v, f.occur_list)
flip += 1
return sat, flip, None