def inference(observations):
use_graphics = False
all_possible_hidden_states = robot.get_all_hidden_states()
all_possible_observed_states = robot.get_all_observed_states()
prior_distribution = robot.initial_distribution()
print('Running Viterbi...')
estimated_states = Viterbi(all_possible_hidden_states,
all_possible_observed_states,
prior_distribution, robot.transition_model,
robot.observation_model, observations)
print('Request handled.')
print('Ready to infer robot behavior.')
return estimated_states
need_to_generate_data = False
num_time_steps = len(hidden_states)
# if no data is loaded, then generate new data
if need_to_generate_data:
num_time_steps = 100
hidden_states, observations = \
generate_data(robot.initial_distribution,
robot.transition_model,
robot.observation_model,
num_time_steps,
make_some_observations_missing)
all_possible_hidden_states = robot.get_all_hidden_states()
all_possible_observed_states = robot.get_all_observed_states()
prior_distribution = robot.initial_distribution()
print 'Running forward-backward...'
marginals = forward_backward(all_possible_hidden_states,
all_possible_observed_states,
prior_distribution,
robot.transition_model,
robot.observation_model,
observations)
print
print "Marginal at time %d:" % (num_time_steps - 1)
if marginals[-1] is not None:
print marginals[-1]
else:
print '*No marginal computed*'
for key in phiLast.keys():
val2 = msgHat[key]
if val2 == 0:
val2 = np.inf
finNode[key] = neglog(phiLast[key]) + val2
minVal, minKey = myDictMin(finNode)
mHat = minVal
tBack = minKey
return mHat, tBack
# %% main program
# project
all_possible_hidden_states = robot.get_all_hidden_states()
all_possible_observed_states = robot.get_all_observed_states()
prior_distribution = robot.initial_distribution()
transition_model = robot.transition_model
observation_model = robot.observation_model
#observations = [(2, 0), (2, 0), (3, 0), (4, 0), (4, 0),
# (6, 0), (6, 1), (5, 0), (6, 0), (6, 2)]
observations = [(1, 6), (4, 6), (4, 7), None, (5, 6),
(6, 5), (6, 6), None, (5, 5), (4, 4)]
## example with class coins
#import classCoinsExample as cc
#all_possible_hidden_states = cc.get_all_hidden_states()
#all_possible_observed_states = cc.get_all_observed_states()
#prior_distribution = cc.initial_distribution()
#transition_model = cc.transition_model
#observation_model = cc.observation_model
def baum_welch(all_possible_hidden_states,
all_possible_observed_states,
observations):
"""
Inputs
------
all_possible_hidden_states: a list of possible hidden states
all_possible_observed_states: a list of possible observed states
observations: a list of observations, one per hidden state
(in this problem, we don't have any missing observations)
Output
------
A transition model and an observation model
"""
### Initialize
transition_model = robot.uniform_transition_model
observation_model = robot.spread_observation_model
initial_state_distribution = robot.initial_distribution()
num_time_steps = len(observations)
convergence_reached = False
while not convergence_reached:
# vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
### YOUR CODE HERE: Estimate marginals & pairwise marginals
pairwise_marginals = [None] * num_time_steps
marginals = [None] * num_time_steps
# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
### Learning: estimate model parameters
# maps states to distributions over possible next states
transition_model_dict = {}
# maps states to distributions over possible observations
observation_model_dict = {}
# initial
initial_state_distribution = robot.Distribution()
# vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
# YOUR CODE HERE: Use the estimated marginals & pairwise marginals to
# estimate the parameters.
# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
new_transition_model = robot.make_transition_model(transition_model_dict)
new_obs_model = robot.make_observation_model(observation_model_dict)
### Check for convergence
if check_convergence(all_possible_hidden_states,
transition_model,
new_transition_model):
convergence_reached = True
transition_model = new_transition_model
observation_model = new_obs_model
return (transition_model,
observation_model,
initial_state_distribution,
marginals)
