def transition_model(state):
# given a hidden state, return the Distribution for the next hidden state
x, y, action = state
next_states = Distribution()
# we can always stay where we are
if action == 'stay':
next_states[(x, y, 'stay')] = .2
else:
next_states[(x, y, 'stay')] = .1
if y > 0: # we can go up
if action == 'stay':
next_states[(x, y-1, 'up')] = .2
if action == 'up':
next_states[(x, y-1, 'up')] = .9
if y < GRID_HEIGHT - 1: # we can go down
if action == 'stay':
next_states[(x, y+1, 'down')] = .2
if action == 'down':
next_states[(x, y+1, 'down')] = .9
if x > 0: # we can go left
if action == 'stay':
next_states[(x-1, y, 'left')] = .2
if action == 'left':
next_states[(x-1, y, 'left')] = .9
if x < GRID_WIDTH - 1: # we can go right
if action == 'stay':
next_states[(x+1, y, 'right')] = .2
if action == 'right':
next_states[(x+1, y, 'right')] = .9
next_states.renormalize()
return next_states
def transition_model(state):
# given a hidden state, return the Distribution for the next hidden state
x, y, action = state
next_states = Distribution()
# we can always stay where we are
if action == 'stay':
next_states[(x, y, 'stay')] = .2
else:
next_states[(x, y, 'stay')] = .1
if y > 0: # we can go up
if action == 'stay':
next_states[(x, y - 1, 'up')] = .2
if action == 'up':
next_states[(x, y - 1, 'up')] = .9
if y < GRID_HEIGHT - 1: # we can go down
if action == 'stay':
next_states[(x, y + 1, 'down')] = .2
if action == 'down':
next_states[(x, y + 1, 'down')] = .9
if x > 0: # we can go left
if action == 'stay':
next_states[(x - 1, y, 'left')] = .2
if action == 'left':
next_states[(x - 1, y, 'left')] = .9
if x < GRID_WIDTH - 1: # we can go right
if action == 'stay':
next_states[(x + 1, y, 'right')] = .2
if action == 'right':
next_states[(x + 1, y, 'right')] = .9
next_states.renormalize()
return next_states
def uniform_transition_model(state):
next_state_distribution = Distribution()
valid_next_states = get_valid_next_states(state)
for next_state in valid_next_states:
next_state_distribution[next_state] += 1
next_state_distribution.renormalize()
return next_state_distribution
def uniform_transition_model(state):
next_state_distribution = Distribution()
valid_next_states = get_valid_next_states(state)
for next_state in valid_next_states:
next_state_distribution[next_state] += 1
next_state_distribution.renormalize()
return next_state_distribution
def compute_marginal(message, node_potential):
marginal = Distribution()
for hidden_state in all_possible_hidden_states:
if hidden_state in message and \
hidden_state in node_potential:
value = message[hidden_state] * node_potential[hidden_state]
if value > 0: # only store entries with nonzero prob.
marginal[hidden_state] = value
marginal.renormalize()
return marginal
def compute_marginal(message, node_potential):
marginal = Distribution()
for hidden_state in all_possible_hidden_states:
if hidden_state in message and \
hidden_state in node_potential:
value = message[hidden_state] * node_potential[hidden_state]
if value > 0: # only store entries with nonzero prob.
marginal[hidden_state] = value
marginal.renormalize()
return marginal
def spread_observation_model(state, radius=1):
# given a hidden state, return the Distribution for its observation
x, y, action = state
observed_states = Distribution()
for x_new in range(x - radius, x + radius + 1):
for y_new in range(y - radius, y + radius + 1):
if x_new >= 0 and x_new <= GRID_WIDTH - 1 and \
y_new >= 0 and y_new <= GRID_HEIGHT - 1:
observed_states[(x_new, y_new)] = 1.
observed_states.renormalize()
return observed_states
def spread_observation_model(state, radius=1):
# given a hidden state, return the Distribution for its observation
x, y, action = state
observed_states = Distribution()
for x_new in range(x - radius, x + radius + 1):
for y_new in range(y - radius, y + radius + 1):
if x_new >= 0 and x_new <= GRID_WIDTH - 1 and \
y_new >= 0 and y_new <= GRID_HEIGHT - 1:
observed_states[(x_new, y_new)] = 1.
observed_states.renormalize()
return observed_states
def discretized_gaussian_observation_model(state, sigma=1):
# given a hidden state, return the Distribution for its observation
x, y, action = state
observed_states = Distribution()
# x_new, y_new = np.meshgrid(range(GRID_WIDTH), range(GRID_HEIGHT))
# values = np.exp( -( (x_new - x)**2 + (y_new -y)**2 )/(2.*sigma) )
for x_new in range(GRID_WIDTH):
for y_new in range(GRID_HEIGHT):
observed_states[(x_new, y_new)] = \
np.exp(-( (x_new - x)**2 + (y_new - y)**2 )/(2.*sigma))
observed_states.renormalize()
return observed_states
def discretized_gaussian_observation_model(state, sigma=1):
# given a hidden state, return the Distribution for its observation
x, y, action = state
observed_states = Distribution()
# x_new, y_new = np.meshgrid(range(GRID_WIDTH), range(GRID_HEIGHT))
# values = np.exp( -( (x_new - x)**2 + (y_new -y)**2 )/(2.*sigma) )
for x_new in range(GRID_WIDTH):
for y_new in range(GRID_HEIGHT):
observed_states[(x_new, y_new)] = \
np.exp(-( (x_new - x)**2 + (y_new - y)**2 )/(2.*sigma))
observed_states.renormalize()
return observed_states
def forward_backward(all_possible_hidden_states, all_possible_observed_states,
prior_distribution, transition_model, observation_model,
observations):
"""
Inputs
------
all_possible_hidden_states: a list of possible hidden states
all_possible_observed_states: a list of possible observed states
prior_distribution: a distribution over states
transition_model: a function that takes a hidden state and returns a
Distribution for the next state
observation_model: a function that takes a hidden state and returns a
Distribution for the observation from that hidden state
observations: a list of observations, one per hidden state
(a missing observation is encoded as None)
Output
------
A list of marginal distributions at each time step; each distribution
should be encoded as a Distribution (see the Distribution class in
robot.py and see how it is used in both robot.py and the function
generate_data() above, and the i-th Distribution should correspond to time
step i
"""
num_time_steps = len(observations)
#-------------------------------------------------------------------------
# Fold observations into singleton potentials
#
phis = [] # phis[n] is the singleton potential for node n
for n in range(num_time_steps):
potential = Distribution()
observed_state = observations[n]
if n == 0:
for hidden_state in prior_distribution:
value = prior_distribution[hidden_state]
if observed_state is not None:
value *= observation_model(hidden_state)[observed_state]
if value > 0: # only store entries with nonzero prob.
potential[hidden_state] = value
else:
for hidden_state in all_possible_hidden_states:
if observed_state is None:
# singleton potential should be identically 1
potential[hidden_state] = 1.
else:
value = observation_model(hidden_state)[observed_state]
if value > 0: # only store entries with nonzero prob.
potential[hidden_state] = value
assert len(potential.keys()) > 0 , \
"Invalid observation at time %d. Maybe you \
forgot the --use-spread-output argument?" %n
phis.append(potential)
# we need not recompute edge potentials since they're given by the
# transition model: phi(x_i, x_j) = transition_model[x_i](x_j),
# where j = i+1
#-------------------------------------------------------------------------
# Forward pass
#
forward_messages = []
# compute message from non-existent node -1 to node 0
message = Distribution()
for hidden_state in all_possible_hidden_states:
message[hidden_state] = 1.
message.renormalize()
forward_messages.append(message)
for n in range(num_time_steps - 1):
# compute message from node n to node n+1
message = Distribution()
## the commented block below is easier to understand but is slow;
## a faster version is below that switches the order of the for loops
## and reduces the number of states that we iterate over
#for next_hidden_state in all_possible_hidden_states:
# value = 0.
# # only loop over hidden states with nonzero singleton potential!
# for hidden_state in phis[n]:
# value += phis[n][hidden_state] * \
# transition_model(hidden_state)[next_hidden_state] * \
# forward_messages[-1][hidden_state]
# if value > 0: # only store entries with nonzero prob.
# message[next_hidden_state] = value
## faster version of the commented block above
# 1. only loop over hidden states with nonzero singleton potential!
for hidden_state in phis[n]:
# 2. only loop over possible next hidden states given current
# hidden state
for next_hidden_state in transition_model(hidden_state):
factor = phis[n][hidden_state] * \
transition_model(hidden_state)[next_hidden_state] * \
forward_messages[-1][hidden_state]
if factor > 0: # only store entries with nonzero prob.
if next_hidden_state in message:
message[next_hidden_state] += factor
else:
message[next_hidden_state] = factor
message.renormalize()
forward_messages.append(message)
#-------------------------------------------------------------------------
# Pre-processing to speed up the backward pass: cache for each hidden
# state what the possible previous hidden states are
#
possible_prev_hidden_states = {}
for hidden_state in all_possible_hidden_states:
for next_hidden_state in transition_model(hidden_state):
if next_hidden_state in possible_prev_hidden_states:
possible_prev_hidden_states[next_hidden_state].add( \
hidden_state)
else:
possible_prev_hidden_states[next_hidden_state] = \
set([hidden_state])
#-------------------------------------------------------------------------
# Backward pass
#
backward_messages = []
# compute message from non-existent node <num_time_steps> to node
# <num_time_steps>-1
message = Distribution()
for hidden_state in all_possible_hidden_states:
message[hidden_state] = 1.
message.renormalize()
backward_messages.append(message)
for n in range(num_time_steps - 2, -1, -1):
# compute message from node n+1 to n
message = Distribution()
## again, I've commented out a block that's easier to understand but
## slow; the faster version is below
#for hidden_state in all_possible_hidden_states:
# value = 0.
# for next_hidden_state in transition_model(hidden_state):
# value += phis[n+1][next_hidden_state] * \
# transition_model(hidden_state)[next_hidden_state] * \
# backward_messages[0][next_hidden_state]
# if value > 0: # only store entries with nonzero prob.
# message[hidden_state] = value
## faster version
# 1. only loop over next hidden states with nonzero potential!
for next_hidden_state in phis[n + 1]:
# 2. only loop over possible previous hidden states
for hidden_state in possible_prev_hidden_states[next_hidden_state]:
factor = phis[n+1][next_hidden_state] * \
transition_model(hidden_state)[next_hidden_state] * \
backward_messages[0][next_hidden_state]
if factor > 0: # only store entries with nonzero prob.
if hidden_state in message:
message[hidden_state] += factor
else:
message[hidden_state] = factor
message.renormalize()
backward_messages.insert(0, message)
#-------------------------------------------------------------------------
# Compute marginals
#
marginals = []
for n in range(num_time_steps):
marginal = Distribution()
for hidden_state in all_possible_hidden_states:
if hidden_state in forward_messages[n] and \
hidden_state in backward_messages[n] and \
hidden_state in phis[n]:
value = forward_messages[n][hidden_state] * \
backward_messages[n][hidden_state] * \
phis[n][hidden_state]
if value > 0: # only store entries with nonzero prob.
marginal[hidden_state] = value
marginal.renormalize()
marginals.append(marginal)
# vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
### YOUR CODE HERE: Estimate marginals & pairwise marginals
pairwise_marginals = [None] * (num_time_steps - 1)
# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
return (marginals, pairwise_marginals)
def forward(all_possible_hidden_states, prior_distribution, transition_model,
observation_model, observations):
"""
Inputs
------
all_possible_hidden_states: a list of possible hidden states
prior_distribution: a distribution over states
transition_model: a function that takes a hidden state and returns a
Distribution for the next state
observation_model: a function that takes a hidden state and returns a
Distribution for the observation from that hidden state
observations: a list of observations, one per hidden state
(a missing observation is encoded as None)
Output
------
This function is a Python generator! It calculates outputs "on demand";
the i-th output should be the marginal distribution for time step i.
"""
num_time_steps = len(observations)
#-------------------------------------------------------------------------
# Forward pass
#
def compute_marginal(message, node_potential):
marginal = Distribution()
for hidden_state in all_possible_hidden_states:
if hidden_state in message and \
hidden_state in node_potential:
value = message[hidden_state] * node_potential[hidden_state]
if value > 0: # only store entries with nonzero prob.
marginal[hidden_state] = value
marginal.renormalize()
return marginal
# compute message from non-existent node -1 to node 0
message = Distribution()
for hidden_state in all_possible_hidden_states:
message[hidden_state] = 1.
message.renormalize()
# compute node potential for time step 0
node_potential = Distribution()
observed_state = observations[0]
for hidden_state in prior_distribution:
value = prior_distribution[hidden_state]
if observed_state is not None:
value *= observation_model(hidden_state)[observed_state]
if value > 0:
node_potential[hidden_state] = value
yield compute_marginal(message, node_potential)
prev_message = message
prev_node_potential = node_potential
for n in range(1, num_time_steps):
message = Distribution()
node_potential = Distribution()
observed_state = observations[n]
# compute message from node n-1 to n and fill in node potential for
# time step n
for hidden_state in all_possible_hidden_states:
# only loop over possible next hidden states given current
# hidden state
for next_hidden_state in transition_model(hidden_state):
factor = prev_node_potential[hidden_state] * \
transition_model(hidden_state)[next_hidden_state] * \
prev_message[hidden_state]
if factor > 0: # only store entries with nonzero prob.
if next_hidden_state in message:
message[next_hidden_state] += factor
else:
message[next_hidden_state] = factor
if observed_state is not None:
value = observation_model(hidden_state)[observed_state]
if value > 0:
node_potential[hidden_state] = value
else:
node_potential[hidden_state] = 1.
message.renormalize()
yield compute_marginal(message, node_potential)
prev_message = message
prev_node_potential = node_potential
def forward(all_possible_hidden_states,
prior_distribution,
transition_model,
observation_model,
observations):
"""
Inputs
------
all_possible_hidden_states: a list of possible hidden states
prior_distribution: a distribution over states
transition_model: a function that takes a hidden state and returns a
Distribution for the next state
observation_model: a function that takes a hidden state and returns a
Distribution for the observation from that hidden state
observations: a list of observations, one per hidden state
(a missing observation is encoded as None)
Output
------
This function is a Python generator! It calculates outputs "on demand";
the i-th output should be the marginal distribution for time step i.
"""
num_time_steps = len(observations)
#-------------------------------------------------------------------------
# Forward pass
#
def compute_marginal(message, node_potential):
marginal = Distribution()
for hidden_state in all_possible_hidden_states:
if hidden_state in message and \
hidden_state in node_potential:
value = message[hidden_state] * node_potential[hidden_state]
if value > 0: # only store entries with nonzero prob.
marginal[hidden_state] = value
marginal.renormalize()
return marginal
# compute message from non-existent node -1 to node 0
message = Distribution()
for hidden_state in all_possible_hidden_states:
message[hidden_state] = 1.
message.renormalize()
# compute node potential for time step 0
node_potential = Distribution()
observed_state = observations[0]
for hidden_state in prior_distribution:
value = prior_distribution[hidden_state]
if observed_state is not None:
value *= observation_model(hidden_state)[observed_state]
if value > 0:
node_potential[hidden_state] = value
yield compute_marginal(message, node_potential)
prev_message = message
prev_node_potential = node_potential
for n in range(1, num_time_steps):
message = Distribution()
node_potential = Distribution()
observed_state = observations[n]
# compute message from node n-1 to n and fill in node potential for
# time step n
for hidden_state in all_possible_hidden_states:
# only loop over possible next hidden states given current
# hidden state
for next_hidden_state in transition_model(hidden_state):
factor = prev_node_potential[hidden_state] * \
transition_model(hidden_state)[next_hidden_state] * \
prev_message[hidden_state]
if factor > 0: # only store entries with nonzero prob.
if next_hidden_state in message:
message[next_hidden_state] += factor
else:
message[next_hidden_state] = factor
if observed_state is not None:
value = observation_model(hidden_state)[observed_state]
if value > 0:
node_potential[hidden_state] = value
else:
node_potential[hidden_state] = 1.
message.renormalize()
yield compute_marginal(message, node_potential)
prev_message = message
prev_node_potential = node_potential
def forward_backward(
all_possible_hidden_states,
all_possible_observed_states,
prior_distribution,
transition_model,
observation_model,
observations,
):
"""
Inputs
------
all_possible_hidden_states: a list of possible hidden states
all_possible_observed_states: a list of possible observed states
prior_distribution: a distribution over states
transition_model: a function that takes a hidden state and returns a
Distribution for the next state
observation_model: a function that takes a hidden state and returns a
Distribution for the observation from that hidden state
observations: a list of observations, one per hidden state
(a missing observation is encoded as None)
Output
------
A list of marginal distributions at each time step; each distribution
should be encoded as a Distribution (see the Distribution class in
robot.py and see how it is used in both robot.py and the function
generate_data() above, and the i-th Distribution should correspond to time
step i
"""
num_time_steps = len(observations)
# -------------------------------------------------------------------------
# Fold observations into singleton potentials
#
phis = [] # phis[n] is the singleton potential for node n
for n in range(num_time_steps):
potential = Distribution()
observed_state = observations[n]
if n == 0:
for hidden_state in prior_distribution:
value = prior_distribution[hidden_state]
if observed_state is not None:
value *= observation_model(hidden_state)[observed_state]
if value > 0: # only store entries with nonzero prob.
potential[hidden_state] = value
else:
for hidden_state in all_possible_hidden_states:
if observed_state is None:
# singleton potential should be identically 1
potential[hidden_state] = 1.0
else:
value = observation_model(hidden_state)[observed_state]
if value > 0: # only store entries with nonzero prob.
potential[hidden_state] = value
assert len(potential.keys()) > 0, (
"Invalid observation at time %d. Maybe you \
forgot the --use-spread-output argument?"
% n
)
phis.append(potential)
# we need not recompute edge potentials since they're given by the
# transition model: phi(x_i, x_j) = transition_model[x_i](x_j),
# where j = i+1
# -------------------------------------------------------------------------
# Forward pass
#
forward_messages = []
# compute message from non-existent node -1 to node 0
message = Distribution()
for hidden_state in all_possible_hidden_states:
message[hidden_state] = 1.0
message.renormalize()
forward_messages.append(message)
for n in range(num_time_steps - 1):
# compute message from node n to node n+1
message = Distribution()
## the commented block below is easier to understand but is slow;
## a faster version is below that switches the order of the for loops
## and reduces the number of states that we iterate over
# for next_hidden_state in all_possible_hidden_states:
# value = 0.
# # only loop over hidden states with nonzero singleton potential!
# for hidden_state in phis[n]:
# value += phis[n][hidden_state] * \
# transition_model(hidden_state)[next_hidden_state] * \
# forward_messages[-1][hidden_state]
# if value > 0: # only store entries with nonzero prob.
# message[next_hidden_state] = value
## faster version of the commented block above
# 1. only loop over hidden states with nonzero singleton potential!
for hidden_state in phis[n]:
# 2. only loop over possible next hidden states given current
# hidden state
for next_hidden_state in transition_model(hidden_state):
factor = (
phis[n][hidden_state]
* transition_model(hidden_state)[next_hidden_state]
* forward_messages[-1][hidden_state]
)
if factor > 0: # only store entries with nonzero prob.
if next_hidden_state in message:
message[next_hidden_state] += factor
else:
message[next_hidden_state] = factor
message.renormalize()
forward_messages.append(message)
# -------------------------------------------------------------------------
# Pre-processing to speed up the backward pass: cache for each hidden
# state what the possible previous hidden states are
#
possible_prev_hidden_states = {}
for hidden_state in all_possible_hidden_states:
for next_hidden_state in transition_model(hidden_state):
if next_hidden_state in possible_prev_hidden_states:
possible_prev_hidden_states[next_hidden_state].add(hidden_state)
else:
possible_prev_hidden_states[next_hidden_state] = set([hidden_state])
# -------------------------------------------------------------------------
# Backward pass
#
backward_messages = []
# compute message from non-existent node <num_time_steps> to node
# <num_time_steps>-1
message = Distribution()
for hidden_state in all_possible_hidden_states:
message[hidden_state] = 1.0
message.renormalize()
backward_messages.append(message)
for n in range(num_time_steps - 2, -1, -1):
# compute message from node n+1 to n
message = Distribution()
## again, I've commented out a block that's easier to understand but
## slow; the faster version is below
# for hidden_state in all_possible_hidden_states:
# value = 0.
# for next_hidden_state in transition_model(hidden_state):
# value += phis[n+1][next_hidden_state] * \
# transition_model(hidden_state)[next_hidden_state] * \
# backward_messages[0][next_hidden_state]
# if value > 0: # only store entries with nonzero prob.
# message[hidden_state] = value
## faster version
# 1. only loop over next hidden states with nonzero potential!
for next_hidden_state in phis[n + 1]:
# 2. only loop over possible previous hidden states
for hidden_state in possible_prev_hidden_states[next_hidden_state]:
factor = (
phis[n + 1][next_hidden_state]
* transition_model(hidden_state)[next_hidden_state]
* backward_messages[0][next_hidden_state]
)
if factor > 0: # only store entries with nonzero prob.
if hidden_state in message:
message[hidden_state] += factor
else:
message[hidden_state] = factor
message.renormalize()
backward_messages.insert(0, message)
# -------------------------------------------------------------------------
# Compute marginals
#
marginals = []
for n in range(num_time_steps):
marginal = Distribution()
for hidden_state in all_possible_hidden_states:
if hidden_state in forward_messages[n] and hidden_state in backward_messages[n] and hidden_state in phis[n]:
value = forward_messages[n][hidden_state] * backward_messages[n][hidden_state] * phis[n][hidden_state]
if value > 0: # only store entries with nonzero prob.
marginal[hidden_state] = value
marginal.renormalize()
marginals.append(marginal)
# vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
### YOUR CODE HERE: Estimate marginals & pairwise marginals
pairwise_marginals = [None] * (num_time_steps - 1)
# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
return (marginals, pairwise_marginals)