def get_all_trajectories_rec_old(pomdp, policy, belief, use_fraction=True):
""" This one first asks for an action, then looks for observations.
for action in actions:
for obs in observations:
"""
if pomdp.is_goal_belief(belief):
# return [[]]
actions = list(policy[belief].keys())
if len(actions) != 1:
msg = 'Expected that a goal belief only had 1 action.'
raise ValueError(msg)
final_action = actions[0]
belief1 = belief2 = belief
ydist = pomdp.get_observations_dist_given_belief(belief1,
use_fraction=use_fraction)
# Just choose first
final_y = list(ydist)[0]
traj = [[dict(action=final_action, obs=final_y,
belief=belief, belief1=belief1,
ydist=ydist, belief2=belief2)]]
return traj
belief = frozendict2(belief)
actions = policy[belief].keys()
trajectories = []
for action in actions:
assert belief in policy
# evolve the belief given action
belief1 = pomdp.evolve(belief, action, use_fraction=use_fraction)
belief1 = frozendict2(belief1)
# what's the dist of observation given the resulting belief
ydist = pomdp.get_observations_dist_given_belief(belief1,
use_fraction=use_fraction)
for y, _ in ydist.items():
belief2 = pomdp.inference(belief=belief1,
observations=y,
use_fraction=use_fraction)
if belief2 == belief:
print('Found fixed point under optimal policy.')
print(' - belief: %s' % belief)
print(' - policy: %s' % policy[belief])
print(' - belief2: %s' % belief)
raise ValueError()
rest = get_all_trajectories_rec(pomdp, policy, belief2,
use_fraction=use_fraction)
for t1 in rest:
traj = [dict(action=action, obs=y,
belief=belief, belief1=belief1,
ydist=ydist, belief2=belief2)] + t1
trajectories.append(traj)
return trajectories
def __init__(self, states_names, decisions):
""" Note that the commands cannot be tuples. """
self.states_names = states_names
self.transitions = {} # (state, obs) -> state
self.commands = {} # (state, obs) -> command
for decision in decisions:
actions = decision['action']
state = dict(decision['state'])
# history = decision['history']
obs = state['last']
del state['last']
state = frozendict2(state)
next_state, real_action = interpret_actions(state, actions)
# warnings.warn('here we are assuming that commands are string')
# assert isinstance(real_action, str)
assert isinstance(next_state, dict)
#
# if isinstance(actions, tuple):
# assert len(actions) == 2, "Otherwise more complicated"
#
# real_action = actions[0]
# assert not isinstance(real_action, tuple)
# rep_action = actions[1]
# next_state = interpret(state, rep_action)
#
# if next_state == state:
# print('Found info action %s, %s, %s -> %s' %
# (state, obs, str(actions), next_state))
# print('But this transition does not work!')
# raise ValueError()
#
# else:
# real_action = actions
# rep_action = None
# next_state = state
if state!=next_state:
# print('Transition %s, %s -> %s' % (state , obs, next_state))
pass
self.transitions[(state, obs)] = next_state
self.commands[(state, obs)] = real_action
self.state0 = frozendict2(dict([s, 0] for s in self.states_names))
self.decisions = decisions
def __init__(self, states_names, decisions):
""" Note that the commands cannot be tuples. """
self.states_names = states_names
self.transitions = {} # (state, obs) -> state
self.commands = {} # (state, obs) -> command
for decision in decisions:
actions = decision['action']
state = dict(decision['state'])
# history = decision['history']
obs = state['last']
del state['last']
state = frozendict2(state)
next_state, real_action = interpret_actions(state, actions)
# warnings.warn('here we are assuming that commands are string')
# assert isinstance(real_action, str)
assert isinstance(next_state, dict)
#
# if isinstance(actions, tuple):
# assert len(actions) == 2, "Otherwise more complicated"
#
# real_action = actions[0]
# assert not isinstance(real_action, tuple)
# rep_action = actions[1]
# next_state = interpret(state, rep_action)
#
# if next_state == state:
# print('Found info action %s, %s, %s -> %s' %
# (state, obs, str(actions), next_state))
# print('But this transition does not work!')
# raise ValueError()
#
# else:
# real_action = actions
# rep_action = None
# next_state = state
if state != next_state:
# print('Transition %s, %s -> %s' % (state , obs, next_state))
pass
self.transitions[(state, obs)] = next_state
self.commands[(state, obs)] = real_action
self.state0 = frozendict2(dict([s, 0] for s in self.states_names))
self.decisions = decisions
def __setitem__(self, attrs, value):
if not isinstance(attrs, dict):
msg = 'Keys to this dictionary must be dicts'
raise ValueError(msg)
# Todo: check all strings
frozen = frozendict2(**attrs)
dict.__setitem__(self, frozen, value)
def f2(stats):
distances = stats['distances']
key = frozendict2(i1=i1, i2=i2, d=d)
if not key in distances:
return None
else:
return distances[key]
def f2(stats):
distances = stats['distances']
key = frozendict2(i1=i1, i2=i2, d=d)
if not key in distances:
return None
else:
return distances[key]
def __setitem__(self, attrs, value):
if not isinstance(attrs, dict):
msg = "Keys to this dictionary must be dicts"
raise ValueError(msg)
# Todo: check all strings
frozen = frozendict2(**attrs)
dict.__setitem__(self, frozen, value)
def inference(self, belief, observations, use_fraction=False):
belief2 = dict()
p_y = self.partition(observations=observations,
state_dist=belief,
use_fraction=use_fraction)
if p_y == 0:
msg = ('Tried to compute inference of p(x) given y but y '
'cannot be generated by any x in p(x)')
# msg += '\nBut p(y|x) is:\n\t%s ' % p
#
raise SimplePOMDP.ObsNotPossible(msg)
for x, p_x in belief.items():
p_y_given_x = self.likelihood(observations=observations, state=x)
p2 = p_x * p_y_given_x / p_y
if p2 > 0:
belief2[x] = p2
belief2 = frozendict2(belief2.items())
s = sum(belief2.values())
if not s == 1.0:
msg = 'Expected %s, got %s. ' % (1.0, s)
msg += '\n %s' % belief2
raise Exception(msg)
return belief2
def get_decisions(trajectories):
""" Returns list of dicts with fields action, state=dict(last=y)
and history """
n = 0
for tr in trajectories:
for i in range(len(tr)):
action = tr[i]['action']
obs = tr[i]['obs']
history = tuple([yh['obs'] for yh in tr[:i]])
n += 1
# print('%03d: u = %s y = %s history %s' % (n, action, obs, len(history)))
d = dict(action=action,
state=frozendict2(last=obs), history=history)
if 'agent_state' in tr[i]:
d['agent_state'] = tr[i]['agent_state']
yield frozendict2(d)
def get_observations_dist(self, state):
robot, intruder = state # @UnusedVariable
# use perfect sensor to detect
_, intruder_detected = list(IntruderPOMDPDec.get_observations_dist(self, state))[0]
rf = self.get_simulated_rangefinder(robot, horizon=self.horizon)
z = (rf, intruder_detected)
return frozendict2({z:1})
def get_decisions(trajectories):
""" Returns list of dicts with fields action, state=dict(last=y)
and history """
n = 0
for tr in trajectories:
for i in range(len(tr)):
action = tr[i]['action']
obs = tr[i]['obs']
history = tuple([yh['obs'] for yh in tr[:i]])
n += 1
# print('%03d: u = %s y = %s history %s' % (n, action, obs, len(history)))
d = dict(action=action,
state=frozendict2(last=obs),
history=history)
if 'agent_state' in tr[i]:
d['agent_state'] = tr[i]['agent_state']
yield frozendict2(d)
def get_all_trajectories_rec(pomdp, policy, belief, use_fraction=True):
""" This one first generates an observation.
for obs in observations:
for action in actions:
"""
if pomdp.is_goal_belief(belief):
return [[]]
belief = frozendict2(belief)
obs_dist = pomdp.get_observations_dist_given_belief(
belief, use_fraction=use_fraction)
trajectories = []
for obs, _ in obs_dist.items():
belief_given_obs = pomdp.inference(belief=belief,
observations=obs,
use_fraction=use_fraction)
assert belief_given_obs in policy
actions = policy[belief_given_obs].keys()
for action in actions:
belief_after_action = pomdp.evolve(belief_given_obs,
action,
use_fraction=use_fraction)
if belief_after_action == belief:
print('Found fixed point under optimal policy.')
print(' - belief: %s' % belief)
print(' - policy: %s' % policy[belief])
print(' - belief2: %s' % belief)
raise ValueError()
rest = get_all_trajectories_rec(pomdp,
policy,
belief_after_action,
use_fraction=use_fraction)
for t1 in rest:
# desc = """
# From belief
# we sampled obs
# then belief1 = belief given obs
# then we chose action
# then belief2 = belief1 evolved with action.
# """
warnings.warn('Need to chenage belief1,2 to meaningful names.')
traj = [
dict(belief=belief,
action=action,
obs=obs,
belief1=belief_given_obs,
ydist=obs_dist,
belief2=belief_after_action)
] + t1
trajectories.append(traj)
return trajectories
def memoizer(f, obj, *args, **kwargs):
if not '__cache' in obj.__dict__:
obj.__dict__['__cache'] = {}
if not f.__name__ in obj.__cache:
obj.__cache[f.__name__] = {}
cache = obj.__cache[f.__name__]
if memoize_instance_show_initial_cache:
if f.__cache_calls == 0 and len(cache) > 0:
# first time we call it
print('For %r I already have a cache of %d calls' %
(f.__name__, len(cache)))
for key in cache:
print('- %s' % str(key))
f.__cache_calls += 1
def get_signature():
args_str = ",".join(str(x) for x in args)
kwargs_str = ",".join('%s=%s' % (k, v) for k, v in kwargs.items())
signature = '%s:%s.%s(%s,%s)' % (obj.__class__.__name__, id(obj), f.__name__,
args_str, kwargs_str)
return signature
key = (args, frozendict2(kwargs))
if key not in cache:
c0 = time.clock()
t0 = time.time()
result = f(obj, *args, **kwargs)
C = time.clock() - c0
T = time.time() - t0
cache[key] = result
if memoize_instance_show_store:
#print('STORE %s = %s' % (cache[key], get_signature()))
perc = 100.0 * f.__cache_hits / f.__cache_calls
n = len(cache)
cache_size = sum([x.__sizeof__() for x in cache.values()])
if cache_size > memoize_instance_show_store_limit:
print('compute time: clock: %sms wall: %sms' % (C * 1000, T * 1000))
print('cache(S): %5d stored %7d calls hits %3.3f%% size %s for %s' %
(n, f.__cache_calls, perc, cache_size, f.__name__))
else:
f.__cache_hits += 1
pass
if f.__cache_calls % memoize_instance_stats_interval == 0:
perc = 100.0 * f.__cache_hits / f.__cache_calls
n = len(cache)
print('cache(R): %5d stored %7d calls hits %3.3f%% for %s' %
(n, f.__cache_calls, perc, f.__name__))
return cache[key]
def memoizer(f, obj, *args, **kwargs):
if not '__cache' in obj.__dict__:
obj.__dict__['__cache'] = {}
if not f.__name__ in obj.__cache:
obj.__cache[f.__name__] = {}
cache = obj.__cache[f.__name__]
if memoize_instance_show_initial_cache:
if f.__cache_calls == 0 and len(cache) > 0:
# first time we call it
print('For %r I already have a cache of %d calls' %
(f.__name__, len(cache)))
for key in cache:
print('- %s' % str(key))
f.__cache_calls += 1
def get_signature():
args_str = ",".join(str(x) for x in args)
kwargs_str = ",".join('%s=%s' % (k, v) for k, v in kwargs.items())
signature = '%s:%s.%s(%s,%s)' % (obj.__class__.__name__, id(obj), f.__name__,
args_str, kwargs_str)
return signature
key = (args, frozendict2(kwargs))
if key not in cache:
c0 = time.clock()
t0 = time.time()
result = f(obj, *args, **kwargs)
C = time.clock() - c0
T = time.time() - t0
cache[key] = result
if memoize_instance_show_store:
# print('STORE %s = %s' % (cache[key], get_signature()))
perc = 100.0 * f.__cache_hits / f.__cache_calls
n = len(cache)
cache_size = sum([x.__sizeof__() for x in cache.values()])
if cache_size > memoize_instance_show_store_limit:
print('compute time: clock: %sms wall: %sms' % (C * 1000, T * 1000))
print('cache(S): %5d stored %7d calls hits %3.3f%% size %s for %s' %
(n, f.__cache_calls, perc, cache_size, f.__name__))
else:
f.__cache_hits += 1
pass
if f.__cache_calls % memoize_instance_stats_interval == 0:
perc = 100.0 * f.__cache_hits / f.__cache_calls
n = len(cache)
print('cache(R): %5d stored %7d calls hits %3.3f%% for %s' %
(n, f.__cache_calls, perc, f.__name__))
return cache[key]
def get_observations_dist(self, state):
robot, intruder = state # @UnusedVariable
# use perfect sensor to detect
_, intruder_detected = list(
IntruderPOMDPDec.get_observations_dist(self, state))[0]
rf = self.get_simulated_rangefinder(robot, horizon=self.horizon)
z = (rf, intruder_detected)
return frozendict2({z: 1})
def make_key(self, f, args, kwargs):
"""
Returns an hashable key.
We don't want to pickle functions, so we
use their name. Might this have problems with decorators?
TODO: investigate
"""
# TODO: use resolving names, so that equivalent calls
# have the same signature
key = (f.__name__, args, frozendict2(kwargs))
return key
def make_key(self, f, args, kwargs):
"""
Returns an hashable key.
We don't want to pickle functions, so we
use their name. Might this have problems with decorators?
TODO: investigate
"""
# TODO: use resolving names, so that equivalent calls
# have the same signature
key = (f.__name__, args, frozendict2(kwargs))
return key
def __getitem__(self, attrs):
if not isinstance(attrs, frozendict2):
check_isinstance(attrs, dict)
attrs = frozendict2(**attrs)
try:
return dict.__getitem__(self, attrs)
except KeyError as e:
msg = str(e)
keys = self.keys()
if keys:
k = most_similar(self.keys(), attrs)
msg += "\n The most similar key is: %s" % str(k)
raise KeyError(msg)
def interpret_actions_one(state, rep_action):
assert isinstance(state, dict), state
assert isinstance(rep_action, str), rep_action
assert '=' in rep_action, 'Invalid rep action %s (state %s)' % (rep_action, state)
comp, value = rep_action.split('=')
value = int(value)
assert comp in state
next_state = dict(**state)
next_state[comp] = value
next_state = frozendict2(next_state)
return next_state
def __getitem__(self, attrs):
if not isinstance(attrs, frozendict2):
check_isinstance(attrs, dict)
attrs = frozendict2(**attrs)
try:
return dict.__getitem__(self, attrs)
except KeyError as e:
msg = str(e)
keys = self.keys()
if keys:
k = most_similar(self.keys(), attrs)
msg += '\n The most similar key is: %s' % str(k)
raise KeyError(msg)
def interpret_actions_one(state, rep_action):
assert isinstance(state, dict), state
assert isinstance(rep_action, str), rep_action
assert '=' in rep_action, 'Invalid rep action %s (state %s)' % (rep_action,
state)
comp, value = rep_action.split('=')
value = int(value)
assert comp in state
next_state = dict(**state)
next_state[comp] = value
next_state = frozendict2(next_state)
return next_state
def get_all_trajectories_rec(pomdp, policy, belief, use_fraction=True):
""" This one first generates an observation.
for obs in observations:
for action in actions:
"""
if pomdp.is_goal_belief(belief):
return [[]]
belief = frozendict2(belief)
obs_dist = pomdp.get_observations_dist_given_belief(belief,
use_fraction=use_fraction)
trajectories = []
for obs, _ in obs_dist.items():
belief_given_obs = pomdp.inference(belief=belief,
observations=obs,
use_fraction=use_fraction)
assert belief_given_obs in policy
actions = policy[belief_given_obs].keys()
for action in actions:
belief_after_action = pomdp.evolve(belief_given_obs,
action, use_fraction=use_fraction)
if belief_after_action == belief:
print('Found fixed point under optimal policy.')
print(' - belief: %s' % belief)
print(' - policy: %s' % policy[belief])
print(' - belief2: %s' % belief)
raise ValueError()
rest = get_all_trajectories_rec(pomdp, policy, belief_after_action,
use_fraction=use_fraction)
for t1 in rest:
# desc = """
# From belief
# we sampled obs
# then belief1 = belief given obs
# then we chose action
# then belief2 = belief1 evolved with action.
# """
warnings.warn('Need to chenage belief1,2 to meaningful names.')
traj = [dict(belief=belief, action=action, obs=obs,
belief1=belief_given_obs,
ydist=obs_dist, belief2=belief_after_action)] + t1
trajectories.append(traj)
return trajectories
def evolve(self, state_dist, action, use_fraction=False):
self.is_state_dist(state_dist)
if use_fraction:
p2 = defaultdict(lambda: Fraction(0))
else:
p2 = defaultdict(lambda: Fraction(0.0))
for s1, p_s1 in state_dist.items():
conditional = self.transition(s1, action)
self.is_state_dist(conditional)
for s2, p_s2 in conditional.items():
if p_s2 * p_s1 > 0:
p2[s2] += p_s2 * p_s1
for s in list(p2):
if p2[s] == 0:
del p2[s]
return frozendict2(p2)
def evolve(self, state_dist, action, use_fraction=False):
self.is_state_dist(state_dist)
if use_fraction:
p2 = defaultdict(lambda:Fraction(0))
else:
p2 = defaultdict(lambda:Fraction(0.0))
for s1, p_s1 in state_dist.items():
conditional = self.transition(s1, action)
self.is_state_dist(conditional)
for s2, p_s2 in conditional.items():
if p_s2 * p_s1 > 0:
p2[s2] += p_s2 * p_s1
for s in list(p2):
if p2[s] == 0:
del p2[s]
return frozendict2(p2)
def add(self, report, report_type, **kwargs):
"""
Adds a report to the collection.
:param report: Promise of a Report object
:param report_type: A string that describes the "type" of the report
:param kwargs: str->str,int,float parameters used for grouping
"""
if not isinstance(report_type, str):
msg = 'Need a string for report_type, got %r.' % describe_value(report_type)
raise ValueError(msg)
from compmake import Promise
if not isinstance(report, Promise):
msg = ('ReportManager is mean to be given Promise objects, '
'which are the output of comp(). Obtained: %s'
% describe_type(report))
raise ValueError(msg)
# check the format is ok
self._check_report_format(report_type, **kwargs)
key = frozendict2(report=report_type, **kwargs)
if key in self.allreports:
msg = 'Already added report for %s' % key
msg += '\n its values is %s' % self.allreports[key]
msg += '\n new value would be %s' % report
raise ValueError(msg)
self.allreports[key] = report
report_type_sane = report_type.replace('_', '')
key_no_report = dict(**key)
del key_no_report['report']
basename = self.html_resources_prefix + report_type_sane
if key_no_report:
basename += '-' + basename_from_key(key_no_report)
dirname = os.path.join(self.outdir, report_type_sane)
filename = os.path.join(dirname, basename)
self.allreports_filename[key] = filename + '.html'
def remove_field(self, field):
""" Returns a copy of this structure, where the given field
is removed from the keys. Throws an error if removing the
field would make the keys not unique. Also throws
an error if the given field is not present in all keys."""
r = self.__class__()
for key in self:
if not field in key:
msg = "Could not find field %r in key %r." % (field, key)
raise ValueError(msg)
key2 = frozendict2(key)
del key2[field]
if key2 in r:
msg = ('Removing field %r from key %r would make it non unique.' %
(field, key))
raise ValueError(msg)
r[key2] = self[key]
return r
def remove_field(self, field):
""" Returns a copy of this structure, where the given field
is removed from the keys. Throws an error if removing the
field would make the keys not unique. Also throws
an error if the given field is not present in all keys."""
r = self.__class__()
for key in self:
if not field in key:
msg = "Could not find field %r in key %r." % (field, key)
raise ValueError(msg)
key2 = frozendict2(key)
del key2[field]
if key2 in r:
msg = "Removing field %r from key %r would make it non unique." % (
field,
key,
)
raise ValueError(msg)
r[key2] = self[key]
return r
def inference(self, belief, observations, use_fraction=False):
belief2 = dict()
p_y = self.partition(observations=observations, state_dist=belief,
use_fraction=use_fraction)
if p_y == 0:
msg = ('Tried to compute inference of p(x) given y but y '
'cannot be generated by any x in p(x)')
# msg += '\nBut p(y|x) is:\n\t%s ' % p
#
raise SimplePOMDP.ObsNotPossible(msg)
for x, p_x in belief.items():
p_y_given_x = self.likelihood(observations=observations, state=x)
p2 = p_x * p_y_given_x / p_y
if p2 > 0:
belief2[x] = p2
belief2 = frozendict2(belief2.items())
s = sum(belief2.values())
if not s == 1.0:
msg = 'Expected %s, got %s. ' % (1.0, s)
msg += '\n %s' % belief2
raise Exception(msg)
return belief2
def get_all_states(self):
state0 = frozendict2(dict([s, 0] for s in self.states_names))
states = set(self.transitions.values())
states.add(state0)
return states
def __contains__(self, attrs):
frozen = frozendict2(**attrs)
return dict.__contains__(self, frozen)
def get(self, report_type, **kwargs):
key = frozendict2(report=report_type, **kwargs)
return self.allreports[key]
def __contains__(self, attrs):
frozen = frozendict2(**attrs)
return dict.__contains__(self, frozen)
def get_all_states(self):
state0 = frozendict2(dict([s, 0] for s in self.states_names))
states = set(self.transitions.values())
states.add(state0)
return states
def add(self, context, report, report_type: str, **kwargs):
"""
Adds a report to the collection.
:param report: Promise of a Report object
:param report_type: A string that describes the "type" of the report
:param kwargs: str->str,int,float parameters used for grouping
"""
from quickapp.compmake_context import CompmakeContext
assert isinstance(context, CompmakeContext)
if not isinstance(report_type, str):
msg = 'Need a string for report_type, got %r.' % describe_value(
report_type)
raise ValueError(msg)
if not isinstance(report, Promise):
msg = ('ReportManager is mean to be given Promise objects, '
'which are the output of comp(). Obtained: %s' %
describe_type(report))
raise ValueError(msg)
# check the format is ok
self._check_report_format(report_type, **kwargs)
key = frozendict2(report=report_type, **kwargs)
if key in self.allreports:
msg = 'Already added report for %s' % key
msg += '\n its values is %s' % self.allreports[key]
msg += '\n new value would be %s' % report
raise ValueError(msg)
self.allreports[key] = report
report_type_sane = report_type.replace('_', '')
key_no_report = dict(**key)
del key_no_report['report']
basename = self.html_resources_prefix + report_type_sane
if key_no_report:
basename += '-' + basename_from_key(key_no_report)
dirname = os.path.join(self.outdir, report_type_sane)
filename = os.path.join(dirname, basename)
self.allreports_filename[key] = filename + '.html'
write_singles = False
if write_singles:
is_root = context.currently_executing == ['root']
if not is_root:
# Add also a single report independent of a global index
# don't create the single report for the ones that are
# defined in the root session
filename_single = os.path.join(dirname, basename) + '_s.html'
# filename_index_dyn = os.path.join(dirname, basename) + '_dyn.html'
report_nid = self.html_resources_prefix + report_type_sane
if key:
report_nid += '-' + basename_from_key(key)
write_job_id = jobid_minus_prefix(context,
report.job_id + '-writes')
# write_report_yaml(report_nid, report_job_id=report.job_id,
# key=key, html_filename=filename_single,
# report_html_indexed=filename_index_dyn)
context.comp(write_report_single,
report=report,
report_nid=report_nid,
report_html=filename_single,
write_pickle=False,
static_dir=self.static_dir,
job_id=write_job_id)
def memoizer(f, *args, **kwargs):
key = (args, frozendict2(kwargs))
if key not in cache:
cache[key] = f(*args, **kwargs)
#print('memoize: %s %d storage' % (obj, len(cache)))
return cache[key]
def get_all_trajectories_rec_old(pomdp, policy, belief, use_fraction=True):
""" This one first asks for an action, then looks for observations.
for action in actions:
for obs in observations:
"""
if pomdp.is_goal_belief(belief):
# return [[]]
actions = list(policy[belief].keys())
if len(actions) != 1:
msg = 'Expected that a goal belief only had 1 action.'
raise ValueError(msg)
final_action = actions[0]
belief1 = belief2 = belief
ydist = pomdp.get_observations_dist_given_belief(
belief1, use_fraction=use_fraction)
# Just choose first
final_y = list(ydist)[0]
traj = [[
dict(action=final_action,
obs=final_y,
belief=belief,
belief1=belief1,
ydist=ydist,
belief2=belief2)
]]
return traj
belief = frozendict2(belief)
actions = policy[belief].keys()
trajectories = []
for action in actions:
assert belief in policy
# evolve the belief given action
belief1 = pomdp.evolve(belief, action, use_fraction=use_fraction)
belief1 = frozendict2(belief1)
# what's the dist of observation given the resulting belief
ydist = pomdp.get_observations_dist_given_belief(
belief1, use_fraction=use_fraction)
for y, _ in ydist.items():
belief2 = pomdp.inference(belief=belief1,
observations=y,
use_fraction=use_fraction)
if belief2 == belief:
print('Found fixed point under optimal policy.')
print(' - belief: %s' % belief)
print(' - policy: %s' % policy[belief])
print(' - belief2: %s' % belief)
raise ValueError()
rest = get_all_trajectories_rec(pomdp,
policy,
belief2,
use_fraction=use_fraction)
for t1 in rest:
traj = [
dict(action=action,
obs=y,
belief=belief,
belief1=belief1,
ydist=ydist,
belief2=belief2)
] + t1
trajectories.append(traj)
return trajectories
def memoizer(f, *args, **kwargs):
key = (args, frozendict2(kwargs))
if key not in cache:
cache[key] = f(*args, **kwargs)
# print('memoize: %s %d storage' % (obj, len(cache)))
return cache[key]
def get_state(self):
return frozendict2(self.state)
def get_state(self):
return frozendict2(self.state)
def add(self, context, report, report_type, **kwargs):
"""
Adds a report to the collection.
:param report: Promise of a Report object
:param report_type: A string that describes the "type" of the report
:param kwargs: str->str,int,float parameters used for grouping
"""
from quickapp.compmake_context import CompmakeContext
assert isinstance(context, CompmakeContext)
if not isinstance(report_type, six.string_types):
msg = 'Need a string for report_type, got %r.' % describe_value(report_type)
raise ValueError(msg)
if not isinstance(report, Promise):
msg = ('ReportManager is mean to be given Promise objects, '
'which are the output of comp(). Obtained: %s'
% describe_type(report))
raise ValueError(msg)
# check the format is ok
self._check_report_format(report_type, **kwargs)
key = frozendict2(report=report_type, **kwargs)
if key in self.allreports:
msg = 'Already added report for %s' % key
msg += '\n its values is %s' % self.allreports[key]
msg += '\n new value would be %s' % report
raise ValueError(msg)
self.allreports[key] = report
report_type_sane = report_type.replace('_', '')
key_no_report = dict(**key)
del key_no_report['report']
basename = self.html_resources_prefix + report_type_sane
if key_no_report:
basename += '-' + basename_from_key(key_no_report)
dirname = os.path.join(self.outdir, report_type_sane)
filename = os.path.join(dirname, basename)
self.allreports_filename[key] = filename + '.html'
write_singles = False
if write_singles:
is_root = context.currently_executing == ['root']
if not is_root:
# Add also a single report independent of a global index
# don't create the single report for the ones that are
# defined in the root session
filename_single = os.path.join(dirname, basename) + '_s.html'
# filename_index_dyn = os.path.join(dirname, basename) + '_dyn.html'
report_nid = self.html_resources_prefix + report_type_sane
if key:
report_nid += '-' + basename_from_key(key)
write_job_id = jobid_minus_prefix(context, report.job_id + '-writes')
# write_report_yaml(report_nid, report_job_id=report.job_id,
# key=key, html_filename=filename_single,
# report_html_indexed=filename_index_dyn)
context.comp(write_report_single,
report=report, report_nid=report_nid,
report_html=filename_single,
write_pickle=False,
static_dir=self.static_dir,
job_id=write_job_id)
def get(self, report_type, **kwargs):
key = frozendict2(report=report_type, **kwargs)
return self.allreports[key]