def get_value(self, state):
# collects feature value distribution
values = []
probs = []
for loc_kv, loc_p in state.distributions[state.keyMap[self.location_feat]].items():
# gets current location
loc = self.world.float2value(self.location_feat, loc_kv[self.location_feat])
loc_freq_feat = get_num_visits_location_key(self.agent, loc)
for loc_freq_kv, loc_freq_p in state.distributions[state.keyMap[loc_freq_feat]].items():
# gets visitation frequency at current location
freq = loc_freq_kv[loc_freq_feat]
if self.inverse:
for time_kv, time_p in state.distributions[state.keyMap[self.time_feat]].items():
# gets current time
time = time_kv[self.time_feat]
values.append(time - freq)
probs.append(loc_p * loc_freq_p * time_p)
else:
values.append(freq)
probs.append(loc_p * loc_freq_p)
# returns weighted average
return np.array(values).dot(np.array(probs)) * self.normalize_factor
def set_reward(self, agent, weight, model=None):
rwd_feat = rewardKey(agent.name)
# compares agent's current location
rwd_tree = {'if': equalRow(self.location_feat, self.all_locations),
None: noChangeMatrix(rwd_feat)}
# get visitation count according to location
for i, loc in enumerate(self.all_locations):
loc_freq_feat = get_num_visits_location_key(agent, loc)
rwd_tree[i] = dynamicsMatrix(rwd_feat, {self.time_feat: 1., loc_freq_feat: -1.}) \
if self.inverse else setToFeatureMatrix(rwd_feat, loc_freq_feat)
agent.setReward(makeTree(rwd_tree), weight * self.normalize_factor, model)
def set_reward(self, agent, weight, model=None):
rwd_feat = rewardKey(agent.name)
# compares agent's current location
rwd_tree = {'if': equalRow(self.location_feat, self.all_locations),
None: noChangeMatrix(rwd_feat)}
# get binary value according to visitation of location
for i, loc in enumerate(self.all_locations):
loc_freq_feat = get_num_visits_location_key(agent, loc)
rwd_tree[i] = {'if': thresholdRow(loc_freq_feat, 1),
True: setToConstantMatrix(rwd_feat, 1),
False: setToConstantMatrix(rwd_feat, 0)}
agent.setReward(makeTree(rwd_tree), weight * self.normalize_factor, model)
def get_locations_frequencies(trajectories, agents, locations):
"""
Gets the visitation frequencies of the agent for each location according to the provided trajectories.
:param list[list[(World, Distribution)]] trajectories: the set of trajectories, containing sequences of state-action pairs.
:param Agent or list[Agent] agents: a list with the agent for each trajectory set whose location frequencies we want to retrieve.
:param list[str] locations: the list of possible world locations.
:rtype: dict[str,float]
:return: the visitation frequencies for each location.
"""
if isinstance(agents, Agent):
agents = [agents] * len(trajectories)
assert len(trajectories) == len(
agents), 'One agent per set of trajectories has to be provided'
data = np.zeros(len(locations))
for i in range(len(trajectories)):
world = trajectories[i][-1][0]
traj_data = []
for loc in locations:
loc_freq_feat = get_num_visits_location_key(agents[i], loc)
traj_data.append(world.getFeature(loc_freq_feat).expectation())
data += traj_data
return dict(zip(locations, data))