def test_closed_system():
c1 = aiger_bv.atom(1, 'c1', signed=False)
a = aiger_bv.atom(1, 'a', signed=False)
dyn = circ2mdp(chain(n=4, state_name='s', action='a'))
dyn <<= (c1 & a).with_output('a').aigbv
c1_coin = coin((1, 8), name='c1')
dyn <<= c1_coin
assert dyn.inputs == {'a'}
assert dyn.outputs == {'s'}
start = {'s_prev': (True, False, False, False, False)}
end = {'s_prev': (False, True, False, False, False)}
assert dyn.prob(start, {'a': (0, )}, end) == 0
assert dyn.prob(start, {'a': (1, )}, end) == c1_coin.prob() == 1 / 8
c2 = aiger_bv.atom(1, 'c2', signed=False)
const_false = aiger_bv.atom(1, 0, signed=False)
state = aiger_bv.atom(5, 's', signed=False)
clip = state == 0b00001
policy = circ2mdp(aiger_bv.ite(clip, const_false, c2).with_output('a'))
policy <<= coin((1, 8), name='c2')
sys = (policy >> dyn).feedback(inputs=['s'],
outputs=['s'],
latches=['s_prev2'],
keep_outputs=True)
assert sys.inputs == set()
assert sys.outputs == {'s'}
def test_biased_coin(k, m):
f = Fraction(*sorted([k, m]))
k, m = f.numerator, f.denominator
for prob in [f, (k, m)]:
k_expr, m_expr = coin((k, m), 'x')
f2 = Fraction(count(k_expr), count(m_expr))
assert f == f2
def sys3():
mdp = sys1()
mdp |= C.circ2mdp(BV.tee(1, {'c': ['c', 'c_next']}))
coin = (~C.coin((1, 2), name='c')).with_output('c')
mdp <<= coin
delay = BV.aig2aigbv(aiger.delay(['c'], [True]))
delay = C.circ2mdp(delay)
return mdp >> delay
def test_mdp_readme():
from aiger_bv import atom
from aiger_coins import circ2mdp
x = atom(3, 'x', signed=False)
y = atom(3, 'y', signed=False)
expr = (x & y).with_output('x&y')
mdp1 = circ2mdp(expr)
dist = aiger_coins.dist((0, 1, 2), name='y')
mdp2 = dist >> mdp1
assert mdp1.inputs == {'x', 'y'}
assert mdp2.inputs == {'x'}
mdp3 = mdp2 | circ2mdp(aiger_bv.identity_gate(3, 'z'))
assert mdp3.inputs == {'x', 'z'}
assert mdp3.outputs == {'x&y', 'z'}
mdp4 = mdp3.feedback(inputs=['z'], outputs=['x&y'], keep_outputs=True)
assert mdp4.inputs == {'x'}
assert mdp4.outputs == {'x&y', 'z'}
action = atom(1, 'action', signed=False)
x_prev = atom(1, 'x_prev', signed=False)
c = atom(1, 'c', signed=False)
x_next = (x_prev & c & action).with_output('x_next')
sys = circ2mdp(x_next).feedback(
keep_outputs=True,
inputs=['x_prev'],
outputs=['x_next'],
initials=[(True, )],
)
sys <<= coin((1, 2), name='c')
assert sys.inputs == {'action'}
assert sys.outputs == {'x_next'}
sys_actions = 3 * [{'action': (True, )}]
states = 3 * [{'x_next': (True, )}]
actions = sys.encode_trc(sys_actions, states)
assert not any(v['c'][0] for v in actions)
sys_actions2, states2 = sys.decode_trc(actions)
assert sys_actions2 == sys_actions
assert states2 == states
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_smoke():
spec = PLTL.atom('a').historically()
spec = BV.aig2aigbv(spec.aig)
spec = C.circ2mdp(spec)
spec <<= C.coin((1, 8), name='c')
spec >>= C.circ2mdp(BV.sink(1, ['c'])) # HACK
bdd, manager, order = to_bdd2(spec, horizon=3)
assert bdd.dag_size == 4
for i in range(order.total_bits*order.horizon):
t = order.time_step(i)
var = manager.var_at_level(i)
action, t2, _ = TIMED_INPUT_MATCHER.match(var).groups()
assert t == int(t2)
decision = action in spec.inputs
assert decision == order.is_decision(i)
def test_find_env_input():
x, c = map(aiger_ptltl.atom, ('x', 'c'))
sys = (x & c).historically().aig
sys = circ2mdp(aiger_bv.aig2aigbv(sys))
sys <<= coin((1, 2), name='c')
assert sys.inputs == {'x'}
assert len(sys.outputs) == 1
out, *_ = sys.outputs
start = end = sys.aigbv.latch2init
action = {'x': (True, )}
coins = sys.find_env_input(start, action, end)
_, lmap = sys.aigbv(inputs={**action, **coins})
assert lmap == end
def test_find_coin_flips():
x, c = map(aiger_ptltl.atom, ('x', 'c'))
sys = (x & c).historically().aig
sys = circ2mdp(aiger_bv.aig2aigbv(sys))
sys <<= coin((1, 2), name='c')
assert sys.inputs == {'x'}
assert len(sys.outputs) == 1
out, *_ = sys.outputs
sys_actions = 3 * [{'x': (True, )}]
states = 3 * [{out: (True, )}]
actions = sys.encode_trc(sys_actions, states)
assert not any(v['c'][0] for v in actions)
sys_actions2, states2 = sys.decode_trc(actions)
assert sys_actions2 == sys_actions
assert states2 == states
def test_biased_coin(k, m):
f = Fraction(*sorted([k, m]))
k, m = f.numerator, f.denominator
for prob in [f, (k, m)]:
assert coin((k, m), 'x').prob() == f
}
def create_sensor(aps):
sensor = BV.aig2aigbv(A.empty())
for name, ap in APS.items():
sensor |= ap.with_output(name).aigbv
return sensor
SENSOR = create_sensor(APS)
DYN = GW.gridworld(8, start=(8, 8), compressed_inputs=True)
SLIP = BV.atom(1, 'c', signed=False).repeat(2) & BV.atom(2, 'a', signed=False)
SLIP = SLIP.with_output('a').aigbv
DYN2 = C.coin((31, 32), 'c') >> C.circ2mdp(DYN << SLIP)
def encode_state(x, y):
x, y = [BV.encode_int(8, 1 << (v - 1), signed=False) for v in (x, y)]
return {'x': tuple(x), 'y': tuple(y)}
def ap_at_state(x, y, in_ascii=False):
state = encode_state(x, y)
obs = SENSOR(state)[0]
if not in_ascii:
return obs
for k, code in {'yellow': 3, 'red': 1}.items():
def sys2():
mdp = sys1()
mdp <<= C.coin((1, 8), name='c')
mdp >>= C.circ2mdp(BV.sink(1, ['c'])) # HACK
return mdp
