def test_sample_smoke():
spec, sys = scenario_reactive()
cspec = concretize(spec, sys, 3)
ctrl = policy(cspec, 3)
actions = ctrl.simulate()
assert len(actions) == 3
assert isinstance(cspec.accepts(actions), bool)
def test_prefix_tree():
spec, sys = scenario_reactive()
encoded = [
[act(True, True), act(True, True),
act(True, True)],
[act(True, True), act(False, True),
act(False, False)],
]
def to_demo(encoded_trc):
io_seq = zip(encoded_trc, sys.aigbv.simulate(encoded_trc))
for inputs, (_, state) in io_seq:
inputs = fn.project(inputs, sys.inputs)
yield inputs, state
# Technical debt where sys_actions and env_actions
# are two different lists.
demos = [list(zip(*to_demo(etrc))) for etrc in encoded]
tree = prefix_tree(sys, demos)
tree.write_dot('foo.dot')
cspec = concretize(spec, sys, 3)
ctrl = fit(cspec, 0.7, bv=True)
lprob = tree.log_likelihood(ctrl, actions_only=True)
assert lprob < 0
assert tree.psat(cspec) == 1 / 2
lprob2 = tree.log_likelihood(ctrl, actions_only=False)
assert lprob2 < lprob
def test_long_horizon():
# TODO: test that more scenarios work with long horizons.
for scenario in [scenario1, scenario_reactive]:
spec, mdp = scenario()
cspec = concretize(spec, mdp, 20)
ctrl = fit(cspec, 0.96)
assert ctrl.psat == pytest.approx(0.96)
def test_flatten():
spec, sys = scenario_reactive()
cspec = concretize(spec, sys, 3)
actions = [act(True, True), act(True, False), act(True, True)]
bits = cspec.flatten(actions)
assert bits == [True, True, True, False, True, True]
assert cspec.unflatten(bits) == actions
def test_psat_monotonicity(scenario):
spec, mdp = scenario()
cspec = concretize(spec, mdp, 3)
prob = 0
for i in range(10):
ctrl = policy(cspec, i)
prev, prob = prob, ctrl.psat
assert prev <= prob
def test_bv_policy():
spec, mdp = scenario_reactive()
cspec = concretize(spec, mdp, 3)
ctrl = fit(cspec, 0.96)
qdd = cspec._as_dfa(qdd=True)
ctrl_bv = BVPolicy(ctrl)
ctrl_bv.prob(qdd.start, {'a': (False, False)})
def test_nx2qdd():
spec, sys = scenario_reactive()
cspec = concretize(spec, sys, 3)
graph, root, _ = spec2graph(cspec, qdd=True)
assert nx.is_directed_acyclic_graph(graph)
assert len(graph.nodes) == 12 + 4
assert len(graph.edges) == 22
for node in graph.nodes:
assert graph.out_degree[node] <= 2
def test_policy_markov_chain():
spec, sys = scenario_reactive()
cspec = concretize(spec, sys, 3)
ctrl = policy(cspec, 3)
adj, _ = ctrl.stochastic_matrix()
assert adj[0, 0] == 1
assert adj[1, 1] == 1
row_sums = adj.sum(axis=1)
assert np.allclose(row_sums, np.ones_like(row_sums))
def test_spec2graph():
spec, sys = scenario_reactive()
cspec = concretize(spec, sys, 3)
graph, root, _ = spec2graph(cspec)
assert nx.is_directed_acyclic_graph(graph)
# BDD size
assert len(graph.nodes) == 10
assert len(graph.edges) == 16
for node in graph.nodes:
assert graph.out_degree[node] <= 2
def test_abstract_trace():
spec, sys = scenario_reactive()
cspec = concretize(spec, sys, 3)
actions = [act(True, True), act(True, False), act(True, True)]
trc = list(cspec.abstract_trace(actions))
for prev, curr in fn.rest(fn.with_prev(trc)):
if prev == curr:
assert prev.node.level == 6
assert prev.node == cspec.manager.false
else:
clvl, cdebt = curr.node.level, curr.debt
plvl, pdebt = prev.node.level, prev.debt
assert (clvl, -cdebt) < (plvl, -pdebt)
def test_concretize():
spec, sys = scenario_reactive()
cspec = concretize(spec, sys, 3)
actions = [act(True, True), act(True, False), act(True, True)]
assert not cspec.accepts(actions)
cspec2 = cspec.toggle(actions)
assert cspec2.accepts(actions)
assert cspec2.imap == cspec.imap
assert set(cspec.imap.keys()) == {'a'}
assert set(cspec.emap.keys()) == {'c'}
def test_reweighted():
spec, sys = scenario_reactive()
# Hack too re-weight coinl
sys2 = C.coin((1, 4), 'c') >> C.MDP(sys.aigbv >> BV.sink(1, ['##valid']))
cspec2 = concretize(spec, sys2, 3)
graph, root, _ = spec2graph(cspec2)
assert nx.is_directed_acyclic_graph(graph)
assert len(graph.nodes) == 12
assert len(graph.edges) == 20
for node in graph.nodes:
assert graph.out_degree[node] <= 2
def test_policy():
spec, sys = scenario_reactive()
cspec = concretize(spec, sys, 3)
ctrls = [policy(cspec)(3), policy(cspec, 3)]
for i, ctrl in enumerate(ctrls):
assert 0 <= ctrl.psat <= 1
assert len(ctrl.ref2action_dist) == 5
assert all(len(v) == 2 for v in ctrl.ref2action_dist.values())
assert all(
sum(v.values()) == pytest.approx(1)
for v in ctrl.ref2action_dist.values())
pctrl = policy(cspec)
# Agent gets monotonically more optimal
psats = [pctrl(x).psat for x in range(10)]
assert all(x >= y for x, y in fn.with_prev(psats, 0))
def test_policy_markov_chain_psat():
spec, sys = scenario_reactive()
cspec = concretize(spec, sys, 3)
adj, _ = fit(cspec, 0.7).stochastic_matrix()
root_vec = sp.sparse.csr_matrix((adj.shape[0], 1))
root_vec[2] = 1
true_vec = sp.sparse.csr_matrix((adj.shape[0], 1))
true_vec[1] = 1
vec = root_vec.T
for _ in range(cspec.order.horizon * cspec.order.total_bits):
vec = vec @ adj
assert (vec @ true_vec).todense() == pytest.approx(0.7)
vec = true_vec
for _ in range(cspec.order.horizon * cspec.order.total_bits):
vec = adj @ vec
assert (root_vec.T @ vec).todense() == pytest.approx(0.7)
def score(spec):
start_time = time.time()
times = {}
print("concretizing spec")
cspec = concretize(spec, mdp, horizon)
print("done spec")
times["build spec"] = time.time() - start_time
if psat is None:
sat_prob = tree.psat(cspec)
else:
sat_prob = psat
start_time = time.time()
print("fitting policy")
ctrl = fit(cspec, sat_prob, bv=True)
print("done fitting")
times["fit"] = time.time() - start_time
start_time = time.time()
print("compute log likelihood of demos")
lprob = tree.log_likelihood(ctrl, actions_only=True)
times["surprise"] = time.time() - start_time
print("\n----------------------------\n")
print(f"BDD size: {cspec.bexpr.dag_size}")
print(f"Controller Size: {ctrl.size}")
print(f"log_prob: {lprob}")
print("\n".join(f"{key}: {val:.2}s" for key, val in times.items()))
print("\n----------------------------\n")
print(times)
return lprob
def test_fit():
spec, sys = scenario_reactive()
cspec = concretize(spec, sys, 3)
assert fit(cspec, 0.7).psat == pytest.approx(0.7)
