def locally_optimize(evaluations, store, goal_expression, domain, functions, negative,
dynamic_streams, visualize, sampling_time=0):
action_plan = store.best_plan
if action_plan is None:
return None
print('\nPostprocessing | Cost: {} | Total Time: {:.3f}'.format(store.best_cost, store.elapsed_time()))
# TODO: postprocess current skeletons as well
task = task_from_domain_problem(domain, get_problem(evaluations, goal_expression, domain, unit_costs=False))
opt_stream_plan, opt_from_obj = recover_opt_stream_plan(evaluations, action_plan, task)
opt_stream_plan += optimistic_process_streams(evaluations_from_stream_plan(evaluations, opt_stream_plan), functions)
opt_action_plan = [(name, tuple(opt_from_obj.get(o, o) for o in args)) for name, args in action_plan]
pddl_plan = [(name, tuple(map(pddl_from_object, args))) for name, args in opt_action_plan]
stream_plan = recover_stream_plan(evaluations, goal_expression, domain,
opt_stream_plan, pddl_plan, negative, unit_costs=False)
stream_plan = get_synthetic_stream_plan(reorder_stream_plan(stream_plan), dynamic_streams)
# TODO: need to make this just streams
opt_evaluations = apply_streams(evaluations, stream_plan)
opt_cost = get_plan_cost(opt_evaluations, opt_action_plan, domain, unit_costs=False)
dump_plans(stream_plan, opt_action_plan, opt_cost)
if visualize:
log_plans(stream_plan, action_plan, None)
create_visualizations(evaluations, stream_plan, None)
store.start_time = time.time()
store.max_cost = store.best_cost
queue = SkeletonQueue(store, evaluations, goal_expression, domain)
queue.new_skeleton(stream_plan, opt_action_plan, opt_cost)
queue.greedily_process()
queue.timed_process(sampling_time)
def solve_skeleton(evaluations={}, opt_solutions=[], max_time=INF, success_cost=0,
max_complexity=INF, reorder=False, visualize=True):
store = SolutionStore(evaluations=evaluations, max_time=max_time, success_cost=success_cost, verbose=True)
skeleton_queue = SkeletonQueue(store, domain=None, disable=True)
for opt_solution in opt_solutions:
# TODO: add and then process later
stream_plan, opt_plan, cost = opt_solution
# stream_plan = replan_with_optimizers(evaluations, stream_plan, domain, externals) or stream_plan
stream_plan = combine_optimizers(evaluations, stream_plan)
# stream_plan = get_synthetic_stream_plan(stream_plan, # evaluations
# [s for s in synthesizers if not s.post_only])
# stream_plan = recover_optimistic_outputs(stream_plan)
if reorder:
# TODO: this blows up memory wise for long stream plans
stream_plan = reorder_stream_plan(store, stream_plan)
num_optimistic = sum(r.optimistic for r in stream_plan) if stream_plan else 0
action_plan = opt_plan.action_plan if is_plan(opt_plan) else opt_plan
print('Stream plan ({}, {}, {:.3f}): {}\nAction plan ({}, {:.3f}): {}'.format(
get_length(stream_plan), num_optimistic, compute_plan_effort(stream_plan), stream_plan,
get_length(action_plan), cost, str_from_plan(action_plan)))
if is_plan(stream_plan) and visualize:
num_iterations = 0
log_plans(stream_plan, action_plan, num_iterations)
#create_visualizations(evaluations, stream_plan, num_iterations)
visualize_stream_plan(stream_plan)
################
# # optimizer_plan = replan_with_optimizers(evaluations, stream_plan, domain, optimizers)
# optimizer_plan = None
# if optimizer_plan is not None:
# # TODO: post process a bound plan
# print('Optimizer plan ({}, {:.3f}): {}'.format(
# get_length(optimizer_plan), compute_plan_effort(optimizer_plan), optimizer_plan))
# skeleton_queue.new_skeleton(optimizer_plan, opt_plan, cost)
# TODO: these are quite different
#skeleton_queue.process(stream_plan, opt_plan, cost, complexity_limit=INF, max_time=INF)
skeleton_queue.process(stream_plan, opt_plan, cost, complexity_limit=max_complexity, max_time=0)
return store.extract_solution()
def solve_abstract(problem,
constraints=PlanConstraints(),
stream_info={},
replan_actions=set(),
unit_costs=False,
success_cost=INF,
max_time=INF,
max_iterations=INF,
max_memory=INF,
initial_complexity=0,
complexity_step=1,
max_complexity=INF,
max_skeletons=INF,
search_sample_ratio=0,
bind=True,
max_failures=0,
unit_efforts=False,
max_effort=INF,
effort_weight=None,
reorder=True,
visualize=False,
verbose=True,
**search_kwargs):
"""
Solves a PDDLStream problem by first planning with optimistic stream outputs and then querying streams
:param problem: a PDDLStream problem
:param constraints: PlanConstraints on the set of legal solutions
:param stream_info: a dictionary from stream name to StreamInfo altering how individual streams are handled
:param replan_actions: the actions declared to induce replanning for the purpose of deferred stream evaluation
:param unit_costs: use unit action costs rather than numeric costs
:param success_cost: the exclusive (strict) upper bound on plan cost to successfully terminate
:param max_time: the maximum runtime
:param max_iterations: the maximum number of search iterations
:param max_memory: the maximum amount of memory
:param initial_complexity: the initial stream complexity limit
:param complexity_step: the increase in the stream complexity limit per iteration
:param max_complexity: the maximum stream complexity limit
:param max_skeletons: the maximum number of plan skeletons (max_skeletons=None indicates not adaptive)
:param search_sample_ratio: the desired ratio of sample time / search time when max_skeletons!=None
:param bind: if True, propagates parameter bindings when max_skeletons=None
:param max_failures: the maximum number of stream failures before switching phases when max_skeletons=None
:param unit_efforts: use unit stream efforts rather than estimated numeric efforts
:param max_effort: the maximum amount of stream effort
:param effort_weight: a multiplier for stream effort compared to action costs
:param reorder: if True, reorder stream plans to minimize the expected sampling overhead
:param visualize: if True, draw the constraint network and stream plan as a graphviz file
:param verbose: if True, print the result of each stream application
:param search_kwargs: keyword args for the search subroutine
:return: a tuple (plan, cost, evaluations) where plan is a sequence of actions
(or None), cost is the cost of the plan (INF if no plan), and evaluations is init expanded
using stream applications
"""
# TODO: select whether to search or sample based on expected success rates
# TODO: no optimizers during search with relaxed_stream_plan
# TODO: locally optimize only after a solution is identified
# TODO: replan with a better search algorithm after feasible
# TODO: change the search algorithm and unit costs based on the best cost
use_skeletons = (max_skeletons is not None)
#assert implies(use_skeletons, search_sample_ratio > 0)
eager_disabled = (effort_weight is None
) # No point if no stream effort biasing
num_iterations = eager_calls = 0
complexity_limit = initial_complexity
evaluations, goal_exp, domain, externals = parse_problem(
problem,
stream_info=stream_info,
constraints=constraints,
unit_costs=unit_costs,
unit_efforts=unit_efforts)
identify_non_producers(externals)
enforce_simultaneous(domain, externals)
compile_fluent_streams(domain, externals)
# TODO: make effort_weight be a function of the current cost
# if (effort_weight is None) and not has_costs(domain):
# effort_weight = 1
load_stream_statistics(externals)
if visualize and not has_pygraphviz():
visualize = False
print(
'Warning, visualize=True requires pygraphviz. Setting visualize=False'
)
if visualize:
reset_visualizations()
streams, functions, negative, optimizers = partition_externals(
externals, verbose=verbose)
eager_externals = list(filter(lambda e: e.info.eager, externals))
positive_externals = streams + functions + optimizers
has_optimizers = bool(optimizers) # TODO: deprecate
assert implies(has_optimizers, use_skeletons)
################
store = SolutionStore(evaluations,
max_time,
success_cost,
verbose,
max_memory=max_memory)
skeleton_queue = SkeletonQueue(store, domain, disable=not has_optimizers)
disabled = set() # Max skeletons after a solution
while (not store.is_terminated()) and (
num_iterations < max_iterations) and (complexity_limit <=
max_complexity):
num_iterations += 1
eager_instantiator = Instantiator(
eager_externals, evaluations) # Only update after an increase?
if eager_disabled:
push_disabled(eager_instantiator, disabled)
if eager_externals:
eager_calls += process_stream_queue(
eager_instantiator,
store,
complexity_limit=complexity_limit,
verbose=verbose)
################
print(
'\nIteration: {} | Complexity: {} | Skeletons: {} | Skeleton Queue: {} | Disabled: {} | Evaluations: {} | '
'Eager Calls: {} | Cost: {:.3f} | Search Time: {:.3f} | Sample Time: {:.3f} | Total Time: {:.3f}'
.format(num_iterations, complexity_limit,
len(skeleton_queue.skeletons), len(skeleton_queue),
len(disabled), len(evaluations), eager_calls,
store.best_cost, store.search_time, store.sample_time,
store.elapsed_time()))
optimistic_solve_fn = get_optimistic_solve_fn(
goal_exp,
domain,
negative,
replan_actions=replan_actions,
reachieve=use_skeletons,
max_cost=min(store.best_cost, constraints.max_cost),
max_effort=max_effort,
effort_weight=effort_weight,
**search_kwargs)
# TODO: just set unit effort for each stream beforehand
if (max_skeletons is None) or (len(skeleton_queue.skeletons) <
max_skeletons):
disabled_axioms = create_disabled_axioms(
skeleton_queue) if has_optimizers else []
if disabled_axioms:
domain.axioms.extend(disabled_axioms)
stream_plan, opt_plan, cost = iterative_plan_streams(
evaluations,
positive_externals,
optimistic_solve_fn,
complexity_limit,
max_effort=max_effort)
for axiom in disabled_axioms:
domain.axioms.remove(axiom)
else:
stream_plan, opt_plan, cost = OptSolution(
INFEASIBLE, INFEASIBLE, INF) # TODO: apply elsewhere
################
#stream_plan = replan_with_optimizers(evaluations, stream_plan, domain, externals) or stream_plan
stream_plan = combine_optimizers(evaluations, stream_plan)
#stream_plan = get_synthetic_stream_plan(stream_plan, # evaluations
# [s for s in synthesizers if not s.post_only])
#stream_plan = recover_optimistic_outputs(stream_plan)
if reorder:
# TODO: this blows up memory wise for long stream plans
stream_plan = reorder_stream_plan(store, stream_plan)
num_optimistic = sum(r.optimistic
for r in stream_plan) if stream_plan else 0
action_plan = opt_plan.action_plan if is_plan(opt_plan) else opt_plan
print('Stream plan ({}, {}, {:.3f}): {}\nAction plan ({}, {:.3f}): {}'.
format(get_length(stream_plan), num_optimistic,
compute_plan_effort(stream_plan), stream_plan,
get_length(action_plan), cost,
str_from_plan(action_plan)))
if is_plan(stream_plan) and visualize:
log_plans(stream_plan, action_plan, num_iterations)
create_visualizations(evaluations, stream_plan, num_iterations)
################
if (stream_plan is INFEASIBLE) and (not eager_instantiator) and (
not skeleton_queue) and (not disabled):
break
if not is_plan(stream_plan):
print('No plan: increasing complexity from {} to {}'.format(
complexity_limit, complexity_limit + complexity_step))
complexity_limit += complexity_step
if not eager_disabled:
reenable_disabled(evaluations, domain, disabled)
#print(stream_plan_complexity(evaluations, stream_plan))
if not use_skeletons:
process_stream_plan(store,
domain,
disabled,
stream_plan,
opt_plan,
cost,
bind=bind,
max_failures=max_failures)
continue
################
#optimizer_plan = replan_with_optimizers(evaluations, stream_plan, domain, optimizers)
optimizer_plan = None
if optimizer_plan is not None:
# TODO: post process a bound plan
print('Optimizer plan ({}, {:.3f}): {}'.format(
get_length(optimizer_plan),
compute_plan_effort(optimizer_plan), optimizer_plan))
skeleton_queue.new_skeleton(optimizer_plan, opt_plan, cost)
allocated_sample_time = (search_sample_ratio * store.search_time) - store.sample_time \
if len(skeleton_queue.skeletons) <= max_skeletons else INF
if skeleton_queue.process(stream_plan, opt_plan, cost,
complexity_limit,
allocated_sample_time) is INFEASIBLE:
break
################
summary = store.export_summary()
summary.update({
'iterations': num_iterations,
'complexity': complexity_limit,
'skeletons': len(skeleton_queue.skeletons),
})
print('Summary: {}'.format(str_from_object(
summary, ndigits=3))) # TODO: return the summary
write_stream_statistics(externals, verbose)
return store.extract_solution()
def constraint_satisfaction(stream_pddl, stream_map, init, terms, stream_info={},
costs=True, max_cost=INF, success_cost=INF, max_time=INF,
unit_efforts=False, max_effort=INF,
max_skeletons=INF, search_sample_ratio=1, verbose=True, **search_args):
# Approaches
# 1) Existential quantification of bindings in goal conditions
# 2) Backtrack useful streams and then schedule. Create arbitrary outputs for not mentioned.
# 3) Construct all useful streams and then associate outputs with bindings
# Useful stream must satisfy at least one fact. How should these assignments be propagated though?
# Make an action that maps each stream result to unbound values?
# TODO: include functions again for cost-sensitive satisfaction
# TODO: convert init into streams to bind certain facts
# TODO: investigate constraint satisfaction techniques for binding instead
# TODO: could also instantiate all possible free parameters even if not useful
# TODO: effort that is a function of the number of output parameters (degrees of freedom)
# TODO: use a CSP solver instead of a planner internally
# TODO: max_iterations?
if not terms:
return {}, 0, init
constraints, negated, functions = partition_facts(set(map(obj_from_existential_expression, terms)))
if not costs:
functions = []
evaluations = evaluations_from_init(init)
goal_facts = set(filter(lambda f: evaluation_from_fact(f) not in evaluations, constraints))
free_parameters = sorted(get_parameters(goal_facts))
print('Parameters:', free_parameters)
externals = parse_stream_pddl(stream_pddl, stream_map, stream_info, unit_efforts=unit_efforts)
stream_results = extract_streams(evaluations, externals, goal_facts)
function_plan = plan_functions(negated + functions, externals)
plan_skeleton = [Assignment(free_parameters)]
cost = get_optimistic_cost(function_plan)
if max_cost < cost:
return None, INF, init
# TODO: detect connected components
# TODO: eagerly evaluate fully bound constraints
# TODO: consider other results if this fails
domain = create_domain(goal_facts)
init_evaluations = evaluations.copy()
store = SolutionStore(evaluations, max_time=max_time, success_cost=success_cost, verbose=verbose)
queue = SkeletonQueue(store, domain, disable=False)
num_iterations = search_time = sample_time = 0
planner = 'ff-astar' # TODO: toggle within reschedule_stream_plan
#last_clusters = set()
#last_success = True
while not store.is_terminated():
num_iterations += 1
start_time = time.time()
print('\nIteration: {} | Skeletons: {} | Skeleton Queue: {} | Evaluations: {} | '
'Cost: {:.3f} | Search Time: {:.3f} | Sample Time: {:.3f} | Total Time: {:.3f}'.format(
num_iterations, len(queue.skeletons), len(queue),
len(evaluations), store.best_cost, search_time, sample_time, store.elapsed_time()))
external_plan = None
if len(queue.skeletons) < max_skeletons:
domain.axioms[:] = create_disabled_axioms(queue, use_parameters=False)
#dominated = are_domainated(last_clusters, clusters)
#last_clusters = clusters
#if last_success or not dominated: # Could also keep a history of results
stream_plan = reschedule_stream_plan(init_evaluations, goal_facts, domain, stream_results,
unique_binding=True, unsatisfiable=True,
max_effort=max_effort, planner=planner, **search_args)
if stream_plan is not None:
external_plan = reorder_stream_plan(combine_optimizers(
init_evaluations, stream_plan + list(function_plan)))
print('Stream plan ({}, {:.3f}): {}'.format(
get_length(external_plan), compute_plan_effort(external_plan), external_plan))
last_success = (external_plan is not None)
search_time += elapsed_time(start_time)
# Once a constraint added for a skeleton, it should only be relaxed
start_time = time.time()
if last_success: # Only works if create_disable_axioms never changes
allocated_sample_time = (search_sample_ratio * search_time) - sample_time
else:
allocated_sample_time = INF
queue.process(external_plan, plan_skeleton, cost=cost,
complexity_limit=INF, max_time=allocated_sample_time)
sample_time += elapsed_time(start_time)
if not last_success and not queue:
break
# TODO: exhaustively compute all plan skeletons and add to queue within the focused algorithm
write_stream_statistics(externals, verbose)
action_plan, cost, facts = revert_solution(store.best_plan, store.best_cost, evaluations)
bindings = bindings_from_plan(plan_skeleton, action_plan)
return bindings, cost, facts
def locally_optimize(evaluations, store, goal_expression, domain, functions,
negative, dynamic_streams):
action_plan = store.best_plan
if action_plan is None:
return None
print('\nPostprocessing') # TODO: postprocess current skeleton as well
task = task_from_domain_problem(
domain,
get_problem(evaluations, goal_expression, domain, unit_costs=False))
plan_instances = get_action_instances(
task, action_plan) + [get_goal_instance(task.goal)]
replace_derived(task, set(), plan_instances)
preimage = filter(lambda a: not a.negated,
plan_preimage(plan_instances, []))
opt_stream_plan = []
opt_from_obj = {}
for stream_result in extract_order(evaluations, preimage):
input_objects = tuple(
opt_from_obj.get(o, o)
for o in stream_result.instance.input_objects)
instance = stream_result.instance.external.get_instance(input_objects)
#assert(not instance.disabled)
instance.disabled = False
opt_results = instance.next_optimistic()
if not opt_results:
continue
opt_result = opt_results[0]
opt_stream_plan.append(opt_result)
for obj, opt in zip(stream_result.output_objects,
opt_result.output_objects):
opt_from_obj[obj] = opt
opt_stream_plan += optimistic_process_streams(
evaluations_from_stream_plan(evaluations, opt_stream_plan), functions)
opt_action_plan = [(name, tuple(opt_from_obj.get(o, o) for o in args))
for name, args in action_plan]
pddl_plan = [(name, map(pddl_from_object, args))
for name, args in opt_action_plan]
stream_plan = recover_stream_plan(evaluations,
goal_expression,
domain,
opt_stream_plan,
pddl_plan,
negative,
unit_costs=False)
stream_plan = reorder_stream_plan(stream_plan)
stream_plan = get_synthetic_stream_plan(stream_plan, dynamic_streams)
print('Stream plan: {}\n'
'Action plan: {}'.format(stream_plan, opt_action_plan))
opt_cost = 0 # TODO: compute this cost for real
store.start_time = time.time()
store.max_cost = store.best_cost
#sampling_time = 10
sampling_time = 0
queue = SkeletonQueue(evaluations, store)
queue.new_skeleton(stream_plan, opt_action_plan, opt_cost)
queue.greedily_process(sampling_time)
def solve_focused(problem,
stream_info={},
action_info={},
synthesizers=[],
max_time=INF,
max_cost=INF,
unit_costs=False,
unit_efforts=False,
effort_weight=None,
eager_layers=1,
search_sampling_ratio=1,
use_skeleton=True,
visualize=False,
verbose=True,
postprocess=False,
**search_kwargs):
"""
Solves a PDDLStream problem by first hypothesizing stream outputs and then determining whether they exist
:param problem: a PDDLStream problem
:param action_info: a dictionary from stream name to ActionInfo for planning and execution
:param stream_info: a dictionary from stream name to StreamInfo altering how individual streams are handled
:param synthesizers: a list of StreamSynthesizer objects
:param max_time: the maximum amount of time to apply streams
:param max_cost: a strict upper bound on plan cost
:param unit_costs: use unit costs rather than numeric costs
:param effort_weight: a multiplier for stream effort compared to action costs
:param eager_layers: the number of eager stream application layers per iteration
:param search_sampling_ratio: the desired ratio of search time / sample time
:param visualize: if True, it draws the constraint network and stream plan as a graphviz file
:param verbose: if True, this prints the result of each stream application
:param postprocess: postprocess the stream plan to find a better solution
:param search_kwargs: keyword args for the search subroutine
:return: a tuple (plan, cost, evaluations) where plan is a sequence of actions
(or None), cost is the cost of the plan, and evaluations is init but expanded
using stream applications
"""
# TODO: return to just using the highest level samplers at the start
# TODO: select whether to search or sample based on expected success rates
solve_stream_plan_fn = relaxed_stream_plan
#solve_stream_plan_fn = relaxed_stream_plan if effort_weight is None else simultaneous_stream_plan
#solve_stream_plan_fn = sequential_stream_plan # simultaneous_stream_plan | sequential_stream_plan
#solve_stream_plan_fn = incremental_stream_plan # incremental_stream_plan | exhaustive_stream_plan
# TODO: warning check if using simultaneous_stream_plan or sequential_stream_plan with non-eager functions
# TODO: no optimizers during search with relaxed_stream_plan
num_iterations = 0
search_time = sample_time = 0
store = SolutionStore(max_time, max_cost,
verbose) # TODO: include other info here?
evaluations, goal_expression, domain, externals = parse_problem(
problem, stream_info)
unit_costs |= not has_costs(domain)
full_action_info = get_action_info(action_info)
load_stream_statistics(externals + synthesizers)
if visualize and not has_pygraphviz():
visualize = False
print(
'Warning, visualize=True requires pygraphviz. Setting visualize=False'
)
if visualize:
reset_visualizations()
eager_externals = list(filter(lambda e: e.info.eager, externals))
streams, functions, negative = partition_externals(externals)
if verbose:
print('Streams: {}\nFunctions: {}\nNegated: {}'.format(
streams, functions, negative))
queue = SkeletonQueue(store, evaluations, goal_expression, domain)
disabled = set()
while not store.is_terminated():
start_time = time.time()
num_iterations += 1
print(
'\nIteration: {} | Queue: {} | Evaluations: {} | Cost: {} '
'| Search Time: {:.3f} | Sample Time: {:.3f} | Total Time: {:.3f}'.
format(num_iterations, len(queue), len(evaluations),
store.best_cost, search_time, sample_time,
store.elapsed_time()))
layered_process_stream_queue(
Instantiator(evaluations, eager_externals), evaluations, store,
eager_layers)
solve_stream_plan = lambda sr: solve_stream_plan_fn(
evaluations,
goal_expression,
domain,
sr,
negative,
max_cost=store.
best_cost, #max_cost=min(store.best_cost, max_cost),
unit_costs=unit_costs,
unit_efforts=unit_efforts,
effort_weight=effort_weight,
**search_kwargs)
#combined_plan, cost = solve_stream_plan(optimistic_process_streams(evaluations, streams + functions))
combined_plan, cost = iterative_solve_stream_plan(
evaluations, streams, functions, solve_stream_plan)
if action_info:
combined_plan = reorder_combined_plan(evaluations, combined_plan,
full_action_info, domain)
print('Combined plan: {}'.format(combined_plan))
stream_plan, action_plan = separate_plan(combined_plan,
full_action_info)
#stream_plan = replan_with_optimizers(evaluations, stream_plan, domain, externals)
stream_plan = combine_optimizers(evaluations, stream_plan)
#stream_plan = get_synthetic_stream_plan(stream_plan, # evaluations
# [s for s in synthesizers if not s.post_only])
stream_plan = reorder_stream_plan(
stream_plan) # TODO: is this redundant when combined_plan?
dump_plans(stream_plan, action_plan, cost)
if (stream_plan is not None) and visualize:
log_plans(stream_plan, action_plan, num_iterations)
create_visualizations(evaluations, stream_plan, num_iterations)
search_time += elapsed_time(start_time)
# TODO: more generally just add the original plan skeleton to the plan
# TODO: cutoff search exploration time at a certain point
start_time = time.time()
allocated_sample_time = search_sampling_ratio * search_time - sample_time
if use_skeleton:
terminate = not process_skeleton_queue(store, queue, stream_plan,
action_plan, cost,
allocated_sample_time)
else:
terminate = not process_disabled(
store, evaluations, domain, disabled, stream_plan, action_plan,
cost, allocated_sample_time, effort_weight is not None)
sample_time += elapsed_time(start_time)
if terminate:
break
if postprocess and (not unit_costs): # and synthesizers
locally_optimize(evaluations, store, goal_expression, domain,
functions, negative, synthesizers, visualize)
write_stream_statistics(externals + synthesizers, verbose)
return revert_solution(store.best_plan, store.best_cost, evaluations)
def solve_focused(problem,
stream_info={},
action_info={},
synthesizers=[],
max_time=INF,
max_cost=INF,
unit_costs=None,
effort_weight=None,
eager_layers=1,
visualize=False,
verbose=True,
postprocess=False,
**search_kwargs):
"""
Solves a PDDLStream problem by first hypothesizing stream outputs and then determining whether they exist
:param problem: a PDDLStream problem
:param action_info: a dictionary from stream name to ActionInfo for planning and execution
:param stream_info: a dictionary from stream name to StreamInfo altering how individual streams are handled
:param max_time: the maximum amount of time to apply streams
:param max_cost: a strict upper bound on plan cost
:param effort_weight: a multiplier for stream effort compared to action costs
:param eager_layers: the number of eager stream application layers per iteration
:param visualize: if True, it draws the constraint network and stream plan as a graphviz file
:param verbose: if True, this prints the result of each stream application
:param search_kwargs: keyword args for the search subroutine
:return: a tuple (plan, cost, evaluations) where plan is a sequence of actions
(or None), cost is the cost of the plan, and evaluations is init but expanded
using stream applications
"""
# TODO: return to just using the highest level samplers at the start
search_sampling_ratio = 1
solve_stream_plan_fn = relaxed_stream_plan if effort_weight is None else simultaneous_stream_plan
# TODO: warning check if using simultaneous_stream_plan or sequential_stream_plan with non-eager functions
num_iterations = 0
search_time = sample_time = 0
store = SolutionStore(max_time, max_cost,
verbose) # TODO: include other info here?
evaluations, goal_expression, domain, stream_name, externals = parse_problem(
problem, stream_info)
compile_to_exogenous(evaluations, domain, externals)
if unit_costs is None:
unit_costs = not has_costs(domain)
full_action_info = get_action_info(action_info)
load_stream_statistics(stream_name, externals + synthesizers)
if visualize:
clear_visualizations()
# TODO: somehow Functions became no longer eager?
eager_externals = list(
filter(lambda e: e.info.eager or type(e) == Function, externals))
streams, functions, negative = partition_externals(externals)
queue = SkeletonQueue(store, evaluations)
# TODO: decide max_sampling_time based on total search_time or likelihood estimates
# TODO: switch to searching if believe chance of search better than sampling
while not store.is_terminated():
num_iterations += 1
print(
'\nIteration: {} | Queue: {} | Evaluations: {} | Cost: {} '
'| Search Time: {:.3f} | Sample Time: {:.3f} | Total Time: {:.3f}'.
format(num_iterations, len(queue), len(evaluations),
store.best_cost, search_time, sample_time,
store.elapsed_time()))
start_time = time.time()
layered_process_stream_queue(
Instantiator(evaluations, eager_externals), evaluations, store,
eager_layers)
solve_stream_plan = lambda sr: solve_stream_plan_fn(
evaluations,
goal_expression,
domain,
sr,
negative,
max_cost=store.best_cost,
#max_cost=min(store.best_cost, max_cost),
unit_costs=unit_costs,
**search_kwargs)
#combined_plan, cost = solve_stream_plan(populate_results(evaluations, streams + functions))
combined_plan, cost = iterative_solve_stream_plan(
evaluations, streams, functions, solve_stream_plan)
if action_info:
combined_plan = reorder_combined_plan(evaluations, combined_plan,
full_action_info, domain)
print('Combined plan: {}'.format(combined_plan))
stream_plan, action_plan = separate_plan(combined_plan,
full_action_info)
stream_plan = reorder_stream_plan(
stream_plan) # TODO: is this strictly redundant?
stream_plan = get_synthetic_stream_plan(stream_plan, synthesizers)
print('Stream plan: {}\n'
'Action plan: {}'.format(stream_plan, action_plan))
search_time += elapsed_time(start_time)
start_time = time.time()
if stream_plan is None:
if not queue:
break
queue.process_until_success()
#queue.fairly_process()
else:
if visualize:
create_visualizations(evaluations, stream_plan, num_iterations)
queue.new_skeleton(stream_plan, action_plan, cost)
queue.greedily_process()
sample_time += elapsed_time(start_time)
start_time = time.time()
queue.timed_process(search_sampling_ratio * search_time - sample_time)
sample_time += elapsed_time(start_time)
if postprocess and (not unit_costs):
locally_optimize(evaluations, store, goal_expression, domain,
functions, negative, synthesizers)
write_stream_statistics(stream_name, externals + synthesizers, verbose)
return revert_solution(store.best_plan, store.best_cost, evaluations)
def solve_focused(problem, constraints=PlanConstraints(),
stream_info={}, action_info={}, synthesizers=[],
max_time=INF, max_iterations=INF, max_skeletons=INF,
unit_costs=False, success_cost=INF,
complexity_step=1,
unit_efforts=False, max_effort=INF, effort_weight=None,
reorder=True, search_sample_ratio=0,
visualize=False, verbose=True, **search_kwargs):
"""
Solves a PDDLStream problem by first hypothesizing stream outputs and then determining whether they exist
:param problem: a PDDLStream problem
:param constraints: PlanConstraints on the set of legal solutions
:param stream_info: a dictionary from stream name to StreamInfo altering how individual streams are handled
:param action_info: a dictionary from stream name to ActionInfo for planning and execution
:param synthesizers: a list of StreamSynthesizer objects
:param max_time: the maximum amount of time to apply streams
:param max_iterations: the maximum number of search iterations
:param max_iterations: the maximum number of plan skeletons to consider
:param unit_costs: use unit action costs rather than numeric costs
:param success_cost: an exclusive (strict) upper bound on plan cost to terminate
:param unit_efforts: use unit stream efforts rather than estimated numeric efforts
:param complexity_step: the increase in the effort limit after each failure
:param max_effort: the maximum amount of effort to consider for streams
:param effort_weight: a multiplier for stream effort compared to action costs
:param reorder: if True, stream plans are reordered to minimize the expected sampling overhead
:param search_sample_ratio: the desired ratio of search time / sample time
:param visualize: if True, it draws the constraint network and stream plan as a graphviz file
:param verbose: if True, this prints the result of each stream application
:param search_kwargs: keyword args for the search subroutine
:return: a tuple (plan, cost, evaluations) where plan is a sequence of actions
(or None), cost is the cost of the plan, and evaluations is init but expanded
using stream applications
"""
# TODO: select whether to search or sample based on expected success rates
# TODO: no optimizers during search with relaxed_stream_plan
num_iterations = search_time = sample_time = eager_calls = 0
complexity_limit = float(INITIAL_COMPLEXITY)
eager_disabled = effort_weight is None # No point if no stream effort biasing
evaluations, goal_exp, domain, externals = parse_problem(
problem, stream_info=stream_info, constraints=constraints,
unit_costs=unit_costs, unit_efforts=unit_efforts)
store = SolutionStore(evaluations, max_time, success_cost, verbose)
full_action_info = get_action_info(action_info)
load_stream_statistics(externals + synthesizers)
if visualize and not has_pygraphviz():
visualize = False
print('Warning, visualize=True requires pygraphviz. Setting visualize=False')
if visualize:
reset_visualizations()
streams, functions, negative = partition_externals(externals, verbose=verbose)
eager_externals = list(filter(lambda e: e.info.eager, externals))
skeleton_queue = SkeletonQueue(store, goal_exp, domain)
disabled = set()
while (not store.is_terminated()) and (num_iterations < max_iterations):
start_time = time.time()
num_iterations += 1
eager_instantiator = Instantiator(eager_externals, evaluations) # Only update after an increase?
if eager_disabled:
push_disabled(eager_instantiator, disabled)
eager_calls += process_stream_queue(eager_instantiator, store, complexity_limit=complexity_limit, verbose=verbose)
print('\nIteration: {} | Complexity: {} | Skeletons: {} | Skeleton Queue: {} | Disabled: {} | Evaluations: {} | '
'Eager Calls: {} | Cost: {:.3f} | Search Time: {:.3f} | Sample Time: {:.3f} | Total Time: {:.3f}'.format(
num_iterations, complexity_limit, len(skeleton_queue.skeletons), len(skeleton_queue), len(disabled),
len(evaluations), eager_calls, store.best_cost, search_time, sample_time, store.elapsed_time()))
optimistic_solve_fn = get_optimistic_solve_fn(goal_exp, domain, negative,
max_cost=min(store.best_cost, constraints.max_cost),
unit_efforts=unit_efforts, max_effort=max_effort,
effort_weight=effort_weight, **search_kwargs)
if (max_skeletons is not None) and (len(skeleton_queue.skeletons) < max_skeletons):
combined_plan, cost = iterative_plan_streams(evaluations, externals, optimistic_solve_fn, complexity_limit,
unit_efforts=unit_efforts, max_effort=max_effort)
else:
combined_plan, cost = INFEASIBLE, INF
if action_info:
combined_plan = reorder_combined_plan(evaluations, combined_plan, full_action_info, domain)
print('Combined plan: {}'.format(combined_plan))
stream_plan, action_plan = separate_plan(combined_plan, full_action_info)
#stream_plan = replan_with_optimizers(evaluations, stream_plan, domain, externals)
stream_plan = combine_optimizers(evaluations, stream_plan)
#stream_plan = get_synthetic_stream_plan(stream_plan, # evaluations
# [s for s in synthesizers if not s.post_only])
if reorder:
stream_plan = reorder_stream_plan(stream_plan) # This may be redundant when using reorder_combined_plan
print('Stream plan ({}, {:.3f}): {}\nAction plan ({}, {:.3f}): {}'.format(
get_length(stream_plan), compute_plan_effort(stream_plan), stream_plan,
get_length(action_plan), cost, str_from_plan(action_plan)))
if is_plan(stream_plan) and visualize:
log_plans(stream_plan, action_plan, num_iterations)
create_visualizations(evaluations, stream_plan, num_iterations)
search_time += elapsed_time(start_time)
if (stream_plan is INFEASIBLE) and (not eager_instantiator) and (not skeleton_queue) and (not disabled):
break
start_time = time.time()
if not is_plan(stream_plan):
complexity_limit += complexity_step
if not eager_disabled:
reenable_disabled(evaluations, domain, disabled)
elif not stream_plan:
store.add_plan(action_plan, cost)
if max_skeletons is None:
process_stream_plan(store, domain, disabled, stream_plan)
else:
allocated_sample_time = (search_sample_ratio * search_time) - sample_time
skeleton_queue.process(stream_plan, action_plan, cost, complexity_limit, allocated_sample_time)
sample_time += elapsed_time(start_time)
write_stream_statistics(externals + synthesizers, verbose)
return store.extract_solution()
def solve_focused(problem,
constraints=PlanConstraints(),
stream_info={},
replan_actions=set(),
max_time=INF,
max_iterations=INF,
initial_complexity=0,
complexity_step=1,
max_skeletons=INF,
bind=True,
max_failures=0,
unit_costs=False,
success_cost=INF,
unit_efforts=False,
max_effort=INF,
effort_weight=None,
reorder=True,
search_sample_ratio=0,
visualize=False,
verbose=True,
**search_kwargs):
"""
Solves a PDDLStream problem by first hypothesizing stream outputs and then determining whether they exist
:param problem: a PDDLStream problem
:param constraints: PlanConstraints on the set of legal solutions
:param stream_info: a dictionary from stream name to StreamInfo altering how individual streams are handled
:param max_time: the maximum amount of time to apply streams
:param max_iterations: the maximum number of search iterations
:param max_skeletons: the maximum number of plan skeletons to consider
:param unit_costs: use unit action costs rather than numeric costs
:param success_cost: an exclusive (strict) upper bound on plan cost to terminate
:param unit_efforts: use unit stream efforts rather than estimated numeric efforts
:param initial_complexity: the initial effort limit
:param complexity_step: the increase in the effort limit after each failure
:param max_effort: the maximum amount of effort to consider for streams
:param effort_weight: a multiplier for stream effort compared to action costs
:param reorder: if True, stream plans are reordered to minimize the expected sampling overhead
:param search_sample_ratio: the desired ratio of search time / sample time
:param visualize: if True, it draws the constraint network and stream plan as a graphviz file
:param verbose: if True, this prints the result of each stream application
:param search_kwargs: keyword args for the search subroutine
:return: a tuple (plan, cost, evaluations) where plan is a sequence of actions
(or None), cost is the cost of the plan, and evaluations is init but expanded
using stream applications
"""
# TODO: select whether to search or sample based on expected success rates
# TODO: no optimizers during search with relaxed_stream_plan
# TODO: locally optimize only after a solution is identified
# TODO: replan with a better search algorithm after feasible
num_iterations = search_time = sample_time = eager_calls = 0
complexity_limit = initial_complexity
# TODO: make effort_weight be a function of the current cost
# TODO: change the search algorithm and unit costs based on the best cost
eager_disabled = effort_weight is None # No point if no stream effort biasing
evaluations, goal_exp, domain, externals = parse_problem(
problem,
stream_info=stream_info,
constraints=constraints,
unit_costs=unit_costs,
unit_efforts=unit_efforts)
store = SolutionStore(evaluations, max_time, success_cost, verbose)
load_stream_statistics(externals)
if visualize and not has_pygraphviz():
visualize = False
print(
'Warning, visualize=True requires pygraphviz. Setting visualize=False'
)
if visualize:
reset_visualizations()
streams, functions, negative, optimizers = partition_externals(
externals, verbose=verbose)
eager_externals = list(filter(lambda e: e.info.eager, externals))
use_skeletons = max_skeletons is not None
has_optimizers = bool(optimizers)
assert implies(has_optimizers, use_skeletons)
skeleton_queue = SkeletonQueue(store, domain, disable=not has_optimizers)
disabled = set() # Max skeletons after a solution
while (not store.is_terminated()) and (num_iterations < max_iterations):
start_time = time.time()
num_iterations += 1
eager_instantiator = Instantiator(
eager_externals, evaluations) # Only update after an increase?
if eager_disabled:
push_disabled(eager_instantiator, disabled)
eager_calls += process_stream_queue(eager_instantiator,
store,
complexity_limit=complexity_limit,
verbose=verbose)
print(
'\nIteration: {} | Complexity: {} | Skeletons: {} | Skeleton Queue: {} | Disabled: {} | Evaluations: {} | '
'Eager Calls: {} | Cost: {:.3f} | Search Time: {:.3f} | Sample Time: {:.3f} | Total Time: {:.3f}'
.format(num_iterations, complexity_limit,
len(skeleton_queue.skeletons), len(skeleton_queue),
len(disabled), len(evaluations), eager_calls,
store.best_cost, search_time, sample_time,
store.elapsed_time()))
optimistic_solve_fn = get_optimistic_solve_fn(
goal_exp,
domain,
negative,
replan_actions=replan_actions,
reachieve=use_skeletons,
max_cost=min(store.best_cost, constraints.max_cost),
max_effort=max_effort,
effort_weight=effort_weight,
**search_kwargs)
# TODO: just set unit effort for each stream beforehand
if (max_skeletons is None) or (len(skeleton_queue.skeletons) <
max_skeletons):
disabled_axioms = create_disabled_axioms(
skeleton_queue) if has_optimizers else []
if disabled_axioms:
domain.axioms.extend(disabled_axioms)
stream_plan, action_plan, cost = iterative_plan_streams(
evaluations, (streams + functions + optimizers),
optimistic_solve_fn,
complexity_limit,
max_effort=max_effort)
for axiom in disabled_axioms:
domain.axioms.remove(axiom)
else:
stream_plan, action_plan, cost = INFEASIBLE, INFEASIBLE, INF
#stream_plan = replan_with_optimizers(evaluations, stream_plan, domain, externals) or stream_plan
stream_plan = combine_optimizers(evaluations, stream_plan)
#stream_plan = get_synthetic_stream_plan(stream_plan, # evaluations
# [s for s in synthesizers if not s.post_only])
if reorder:
stream_plan = reorder_stream_plan(
stream_plan
) # This may be redundant when using reorder_combined_plan
num_optimistic = sum(r.optimistic
for r in stream_plan) if stream_plan else 0
print('Stream plan ({}, {}, {:.3f}): {}\nAction plan ({}, {:.3f}): {}'.
format(get_length(stream_plan), num_optimistic,
compute_plan_effort(stream_plan), stream_plan,
get_length(action_plan), cost,
str_from_plan(action_plan)))
if is_plan(stream_plan) and visualize:
log_plans(stream_plan, action_plan, num_iterations)
create_visualizations(evaluations, stream_plan, num_iterations)
search_time += elapsed_time(start_time)
if (stream_plan is INFEASIBLE) and (not eager_instantiator) and (
not skeleton_queue) and (not disabled):
break
start_time = time.time()
if not is_plan(stream_plan):
complexity_limit += complexity_step
if not eager_disabled:
reenable_disabled(evaluations, disabled)
#print(stream_plan_complexity(evaluations, stream_plan))
if use_skeletons:
#optimizer_plan = replan_with_optimizers(evaluations, stream_plan, domain, optimizers)
optimizer_plan = None
if optimizer_plan is not None:
# TODO: post process a bound plan
print('Optimizer plan ({}, {:.3f}): {}'.format(
get_length(optimizer_plan),
compute_plan_effort(optimizer_plan), optimizer_plan))
skeleton_queue.new_skeleton(optimizer_plan, action_plan, cost)
allocated_sample_time = (search_sample_ratio * search_time) - sample_time \
if len(skeleton_queue.skeletons) <= max_skeletons else INF
skeleton_queue.process(stream_plan, action_plan, cost,
complexity_limit, allocated_sample_time)
else:
process_stream_plan(store,
domain,
disabled,
stream_plan,
action_plan,
cost,
bind=bind,
max_failures=max_failures)
sample_time += elapsed_time(start_time)
write_stream_statistics(externals, verbose)
return store.extract_solution()