def simulate(hmm,horizon):
"""returns a pair of (state sequence, observation sequence) of length horizon.
for each time t, the agent is in state_sequence[t] and
observes observation_sequence[t]
"""
state = sample_one(hmm.indist)
obsseq=[]
stateseq=[]
for time in range(horizon):
stateseq.append(state)
newobs = {obs:sample_one({0:1-hmm.pobs[obs][state],1:hmm.pobs[obs][state]})
for obs in hmm.obsvars}
obsseq.append(newobs)
state = sample_one(hmm.trans[state])
return stateseq,obsseq
def simobs(hmm,stateseq):
"""returns observation sequence for the state sequence"""
obsseq=[]
for state in stateseq:
newobs = {obs:sample_one({0:1-hmm.pobs[obs][state],1:hmm.pobs[obs][state]})
for obs in hmm.obsvars}
obsseq.append(newobs)
return obsseq
def query(self,
qvar,
obs={},
number_samples=1000,
burn_in=100,
sample_order=None):
"""computes P(qvar|obs) where
qvar is a variable.
obs is a variable:value dictionary.
sample_order is a list of non-observed variables in order.
"""
counts = {val: 0 for val in qvar.domain}
if sample_order is not None:
variables = sample_order
else:
variables = [v for v in self.bn.variables if v not in obs]
var_to_factors = {v: set() for v in self.bn.variables}
for fac in self.bn.factors:
for var in fac.variables:
var_to_factors[var].add(fac)
sample = {var: random.choice(var.domain) for var in variables}
self.display(2, "Sample:", sample)
sample.update(obs)
for i in range(burn_in + number_samples):
if sample_order == None:
random.shuffle(variables)
for var in variables:
# get probability distribution of var given its neighbours
vardist = {val: 1 for val in var.domain}
for val in var.domain:
sample[var] = val
for fac in var_to_factors[var]: # Markov blanket
vardist[val] *= fac.get_value(sample)
sample[var] = sample_one(vardist)
if i >= burn_in:
counts[sample[qvar]] += 1
tot = sum(counts.values())
return counts, {c: v / tot for (c, v) in counts.items()}
def advance(self):
"""advance to the next time.
This assumes that all of the weights are 1."""
self.particles = [
sample_one(self.hmm.trans[st]) for st in self.particles
]
def __init__(self, hmm, number_particles=1000):
self.hmm = hmm
self.particles = [
sample_one(hmm.indist) for i in range(number_particles)
]
self.weights = [1 for i in range(number_particles)]
