def experiments():
base = dict(
domain_dir="visitall-opt11-strips",
domain="domain.pddl",
test_domain="domain.pddl",
feature_namer=visitall_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
distinguish_goals=True,
)
exps = dict()
exps["small"] = update_dict(
base,
instances=[
'problem03-full.pddl',
# 'problem04-full.pddl',
# 'problem05-full.pddl',
],
test_instances=[],
test_policy_instances=all_test_instances(),
max_concept_size=8,
distance_feature_max_complexity=8,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
)
exps["debug"] = update_dict(
base,
instances=[
# 'problem02-full.pddl',
# 'problem03-full.pddl',
"tests.pddl",
],
test_instances=[
# 'problem03-full.pddl',
# 'problem04-full.pddl',
],
# test_policy_instances=all_test_instances(),
# max_concept_size=8,
# distance_feature_max_complexity=8,
# cond_feature_max_complexity=8,
use_equivalence_classes=True,
# use_feature_dominance=True,
# use_incremental_refinement=True,
print_denotations=True,
feature_generator=debug_features,
)
return exps
def experiments():
base = dict(
domain_dir="log_one_truck",
domain="domain-onetr_without_restrictions.pddl",
test_domain="domain-onetr_without_restrictions.pddl",
feature_namer=logistics_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
distinguish_goals=True,
)
exps = dict()
# Goal: arbitrary logistics goal
exps["small"] = update_dict(
base,
instances=['probLOGISTICS-1-1.pddl'],
# test_instances=["prob{:02d}.pddl".format(i) for i in range(2, 5)],
test_instances=[],
test_policy_instances=['probLOGISTICS-4-2.pddl'], # all_instances(),
distance_feature_max_complexity=8,
max_concept_size=8,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
)
return exps
def experiments():
base = dict(
domain_dir="floortile-opt11-strips",
domain="domain.pddl",
test_domain="domain.pddl",
feature_namer=floortile_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
distinguish_goals=True,
)
exps = dict()
exps["small"] = update_dict(
base,
instances=["training1.pddl"],
# instances=["testing.pddl"],
test_instances=[],
test_policy_instances=["opt-p01-002.pddl"],
max_concept_size=8,
distance_feature_max_complexity=8,
parameter_generator=None,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
)
return exps
def experiments():
base = dict(
domain_dir="satellite",
domain="domain.pddl",
test_domain="domain.pddl",
feature_namer=satellite_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
)
exps = dict()
exps["small"] = update_dict(
base,
instances=[
'p01-pfile1.pddl',
'p02-pfile2.pddl',
],
test_instances=[],
test_policy_instances=all_test_instances(),
max_concept_size=8,
parameter_generator=None,
use_equivalence_classes=True,
# use_feature_dominance=True,
# use_incremental_refinement=True,
)
return exps
def setup_workspace(config, sample, k):
""" Create the workspace and update the config for the incremental search with the given parameters """
# Create a subdir, e.g. inc_k5_s20, to place all the data for this learning attempt
subdir = os.path.join(config.experiment_dir, "inc_k{}_s{}".format(k, sample.num_states()))
os.makedirs(subdir, exist_ok=True)
save(subdir, dict(sample=sample))
log_sampled_states(sample, os.path.join(subdir, "sample.txt"))
logging.info("Search for abstraction: |S|={}, k=|{}|".format(sample.num_states(), k))
logging.info("Data stored in {}".format(subdir))
subconfig = update_dict(config.to_dict(), max_concept_size=k, experiment_dir=subdir)
return subconfig
def experiments():
base = dict(
domain_dir="reward",
domain="domain.pddl",
test_domain="domain.pddl",
feature_namer=reward_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
distinguish_goals=True,
)
exps = dict()
exps["small"] = update_dict(
base,
instances=["training_5x5.pddl"],
# instances=["instance_5.pddl", "instance_4_blocked.pddl"],
test_instances=[],
test_policy_instances=all_test_instances(),
max_concept_size=8,
distance_feature_max_complexity=8,
parameter_generator=no_parameter,
# parameter_generator=None
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
)
exps["debug"] = update_dict(exps["small"],
feature_generator=debug_features)
# One reason for overfitting: in a 3x3 grid, with 2 booleans per dimension you can perfectly represent any position
return exps
def experiments():
base = dict(
domain_dir="childsnack-opt14-strips",
domain="domain.pddl",
test_domain="domain.pddl",
feature_namer=childsnack_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
)
exps = dict()
exps["small"] = update_dict(
base,
# instances=['sample{:02d}.pddl'.format(i) for i in range(1, 5)],
instances=[
# 'sample_mini.pddl',
'sample01.pddl',
'sample02.pddl',
# 'sample03.pddl',
# 'child-snack_pfile01-2.pddl' # STATE SPACE TOO LARGE
],
test_instances=[
# 'child-snack_pfile01-2.pddl',
],
test_policy_instances=all_test_instances(),
# create_goal_features_automatically=True,
max_concept_size=10,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
)
return exps
def experiments():
base = dict(
domain_dir="gridworld",
domain="domain_strips.pddl",
test_domain="domain_strips.pddl",
feature_namer=gridworld_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
)
exps = dict()
# 3x3 overfits: in a 3x3 grid, with 2 booleans per dimension you can represent the position
exps["5x5"] = update_dict(
base,
instances=[
"instance_strips_5.pddl",
"instance_strips_5_not_corner.pddl",
"instance_strips_5_not_corner_2.pddl",
],
test_instances=[],
test_policy_instances=[],
max_concept_size=8,
distance_feature_max_complexity=8,
# parameter_generator=gripper_parameters, # Works also, but no real advantage
parameter_generator=add_domain_parameters_strips,
use_equivalence_classes=True,
# use_feature_dominance=True,
# use_incremental_refinement=True,
# feature_generator=generate_features_1,
)
return exps
def experiments():
base = dict(
domain_dir="gripper",
domain="domain.pddl",
test_domain="domain.pddl",
feature_namer=gripper_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
distinguish_goals=True,
)
exps = dict()
exps["small"] = update_dict(
base,
# instances=["sample-2balls.pddl", "sample-small.pddl"],
instances=["prob01.pddl"],
# test_instances=[f"prob{i:02d}.pddl" for i in range(3, 11)],
test_instances=[],
test_policy_instances=[f"prob{i:02d}.pddl" for i in range(3, 21)],
max_concept_size=8,
distance_feature_max_complexity=8,
# parameter_generator=gripper_parameters, # Works also, but no real advantage
parameter_generator=None,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
)
return exps
def experiments():
base = dict(
domain_dir="delivery",
domain="domain.pddl",
test_domain="domain.pddl",
feature_namer=delivery_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
distinguish_goals=True,
)
exps = dict()
exps["small"] = update_dict(
base,
instances=[
'instance_3_3_0.pddl', # Use one small instance with three packages
'instance_4_2_0.pddl', # And a slightly larger one with two packages
# 'instance_5_0.pddl',
],
test_policy_instances=all_test_instances(),
max_concept_size=8,
distance_feature_max_complexity=14,
# feature_generator=debug_features,
# d2l_policy=debug_policy,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
# print_denotations=True,
)
return exps
def experiments():
base = dict(
domain_dir="barman-opt11-strips",
domain="domain.pddl",
test_domain="domain.pddl",
feature_namer=barman_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
)
exps = dict()
exps["small"] = update_dict(
base,
instances=[
'sample01.pddl',
],
test_instances=[
# 'child-snack_pfile01-2.pddl',
],
test_policy_instances=all_test_instances(),
max_concept_size=8,
distance_feature_max_complexity=8,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
)
return exps
def experiments():
base = dict(
domain_dir="blocks",
feature_namer=blocksworld_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
test_instances=[],
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
distinguish_goals=True,
)
exps = dict()
strips_base = update_dict(
base,
domain="domain.pddl",
test_domain="domain.pddl",
)
fn_base = update_dict(
base,
domain="domain_fstrips.pddl",
test_domain="domain_fstrips.pddl",
)
strips_atomic_base = update_dict(
base,
domain="domain_atomic_move.pddl",
test_domain="domain_atomic_move.pddl",
)
exps["clear"] = update_dict(
strips_base,
instances=["training_clear_5.pddl"],
test_policy_instances=all_clear_test_instancess(),
max_concept_size=8,
parameter_generator=blocksworld_parameters_for_clear,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
# feature_generator=debug_features_clear,
)
exps["clear_fn"] = update_dict(
fn_base,
instances=["training_clear_5_fs.pddl"],
test_policy_instances=[],
max_concept_size=8,
parameter_generator=blocksworld_parameters_for_clear,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
)
exps["on"] = update_dict(
strips_base,
instances=[
"training_on_5.pddl",
],
test_policy_instances=all_on_test_instancess(),
max_concept_size=8,
distance_feature_max_complexity=8,
parameter_generator=blocksworld_parameters_for_on,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
)
exps["on_fn"] = update_dict(
fn_base,
instances=["training_on_5_fs.pddl"],
test_policy_instances=[],
max_concept_size=8,
parameter_generator=blocksworld_parameters_for_on,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
)
exps["4op_5"] = update_dict(
strips_base,
instances=["probBLOCKS-5-0.pddl"],
test_instances=[],
test_policy_instances=all_4op_test_instances(),
max_concept_size=12,
use_incremental_refinement=True,
use_equivalence_classes=True,
use_feature_dominance=False,
)
exps["4op_debug"] = update_dict(
exps["4op_5"],
instances=["probBLOCKS-4-0.pddl"],
# d2l_policy=debug_policy_4op,
feature_generator=debug_features_4op,
use_incremental_refinement=False,
use_equivalence_classes=True,
use_feature_dominance=False,
)
exps["all_fn_5"] = update_dict(
fn_base,
instances=[
"training_arbitrary_5_fs.pddl",
# "training_singletower_5_fs.pddl",
],
test_policy_instances=[],
max_concept_size=10,
use_incremental_refinement=False,
use_equivalence_classes=True,
use_feature_dominance=False,
# feature_generator=fs_debug_features,
# d2l_policy=fs_debug_policy
)
exps["all_at_5"] = update_dict(
strips_atomic_base,
instances=[
"training_arbitrary_5_atomic.pddl",
"training_arbitrary_5_atomic_tower.pddl"
],
test_policy_instances=[
"training_arbitrary_5_atomic.pddl",
] + [
f"test_atomic_{n}_{i}.pddl" for n in range(10, 31)
for i in range(0, 5)
],
max_concept_size=8,
distance_feature_max_complexity=8,
# feature_generator=debug_features_at,
use_incremental_refinement=True,
use_equivalence_classes=True,
use_feature_dominance=False,
)
# Using incompletely-specified goals
exps["all_at_5_inc"] = update_dict(
exps["all_at_5"],
instances=[
"training_arbitrary_5_atomic_incomplete.pddl",
],
)
exps["all_at_testing"] = update_dict(
strips_atomic_base,
instances=[
"training_arbitrary_5_atomic.pddl",
],
test_policy_instances=[
"training_arbitrary_5_atomic.pddl",
"testing_arbitrary_10_atomic.pddl",
"testing_arbitrary_10_1_atomic.pddl",
"testing_arbitrary_17-0_atomic.pddl",
"testing_arbitrary_17-1_atomic.pddl",
],
feature_generator=debug_features_at2,
use_incremental_refinement=False,
use_equivalence_classes=True,
use_feature_dominance=False,
)
return exps
def experiments():
base = dict(
domain_dir="miconic",
# domain="domain.pddl",
# test_domain="domain.pddl",
# Without the fix, the "board" action allows to board passengers that are not on the floor anymore!
test_domain="domain-with-fix.pddl",
domain="domain-with-fix.pddl",
feature_namer=miconic_names,
pipeline="d2l_pipeline",
maxsat_encoding="d2l",
num_states="all",
concept_generator=None,
parameter_generator=None,
v_slack=2,
# concept_generation_timeout=120, # in seconds
maxsat_timeout=None,
distinguish_goals=True,
)
exps = dict()
exps["small"] = update_dict(
base,
instances=[
# 's2-0.pddl',
# 's2-1.pddl',
# 's2-2.pddl',
# 's2-3.pddl',
# 's3-0.pddl',
's4-0.pddl',
'training2.pddl',
],
test_instances=[],
test_policy_instances=all_test_instances(),
max_concept_size=8,
distance_feature_max_complexity=8,
parameter_generator=None,
use_equivalence_classes=True,
# use_feature_dominance=True,
use_incremental_refinement=True,
decreasing_transitions_must_be_good=True,
)
exps["small2"] = update_dict(
exps["small"],
instances=[
's3-0.pddl',
'training2.pddl',
],
use_incremental_refinement=False,
)
exps["debug"] = update_dict(
exps["small"],
instances=[
# 'debug.pddl',
's4-0.pddl',
'training2.pddl',
],
feature_generator=debug_features,
# d2l_policy=debug_policy,
print_denotations=True,
use_incremental_refinement=False,
test_policy_instances=all_test_instances(),
# test_policy_instances=[
# 's4-0.pddl',
# ]
)
return exps
def experiment(experiment_name=None):
domain = "domain.pddl"
domain_dir = "visitall"
benchmark_dir = BENCHMARK_DIR
# domain_dir = "gripper"
problem02full = dict(
lp_max_weight=10,
benchmark_dir=benchmark_dir,
instances=["problem02-full.pddl"], #, "problem03-full.pddl"],
test_instances=["problem04-full.pddl"],
test_domain=domain,
distance_feature_max_complexity=10,
num_states=500,
num_sampled_states=None,
random_seed=12,
max_concept_size=8,
max_concept_grammar_iterations=3,
concept_generator=None,
parameter_generator=add_domain_parameters,
feature_namer=feature_namer,
)
problem03full = dict(
lp_max_weight=2,
benchmark_dir=benchmark_dir,
instances=["problem02-full.pddl", "problem03-full.pddl"],
test_instances=[
"problem04-full.pddl",
"problem05-full.pddl",
"problem06-full.pddl",
"problem07-full.pddl",
"problem08-full.pddl",
"problem09-full.pddl",
"problem10-full.pddl",
"problem11-full.pddl",
],
test_domain=domain,
distance_feature_max_complexity=10,
num_states=1000,
num_sampled_states=None,
random_seed=12,
max_concept_size=8,
max_concept_grammar_iterations=3,
concept_generator=None,
parameter_generator=add_domain_parameters,
feature_namer=feature_namer,
)
# Incremental version
# Experiment used in the paper
problem03full_incremental = dict(
benchmark_dir=BENCHMARK_DIR,
lp_max_weight=5,
experiment_class=IncrementalExperiment,
test_domain=domain,
instances=[
'training01.pddl',
'training02.pddl',
'training03.pddl',
'training04.pddl',
'training05.pddl',
'training06.pddl',
'training09.pddl',
'training10.pddl',
'training11.pddl',
],
test_instances=[
"problem02-full.pddl",
"problem03-full.pddl",
"problem04-full.pddl",
"problem05-full.pddl",
"problem06-full.pddl",
"problem07-full.pddl",
"problem08-full.pddl",
"problem09-full.pddl",
"problem10-full.pddl",
"problem11-full.pddl",
'p-1-10.pddl',
'p-1-11.pddl',
'p-1-12.pddl',
'p-1-13.pddl',
'p-1-14.pddl',
'p-1-15.pddl',
'p-1-16.pddl',
'p-1-17.pddl',
'p-1-18.pddl',
'p-1-5.pddl',
'p-1-6.pddl',
'p-1-7.pddl',
'p-1-8.pddl',
'p-1-9.pddl',
],
num_states=10000,
initial_sample_size=100,
distance_feature_max_complexity=5,
max_concept_grammar_iterations=10,
initial_concept_bound=8,
max_concept_bound=12,
concept_bound_step=2,
batch_refinement_size=50,
clean_workspace=False,
parameter_generator=add_domain_parameters,
feature_namer=feature_namer,
)
guillem = update_dict(problem03full_incremental,
feature_generator=build_expected_features)
visitall_incremental = dict(
benchmark_dir=BENCHMARK_DIR,
lp_max_weight=100,
experiment_class=IncrementalExperiment,
test_domain=domain,
instances=[
'training01.pddl',
'training02.pddl',
'training03.pddl',
'training04.pddl',
'training05.pddl',
'training06.pddl',
'training09.pddl',
'training10.pddl',
'training11.pddl',
],
test_instances=[
"problem02-full.pddl",
"problem02-half.pddl",
"problem03-full.pddl",
"problem03-half.pddl",
'problem02-full.pddl',
'problem02-half.pddl',
'problem03-full.pddl',
'problem03-half.pddl',
'problem04-full.pddl',
'problem04-half.pddl',
'problem05-full.pddl',
'problem05-half.pddl',
'problem06-full.pddl',
'problem06-half.pddl',
'problem07-full.pddl',
'problem07-half.pddl',
'problem08-full.pddl',
'problem08-half.pddl',
'problem09-full.pddl',
'problem09-half.pddl',
'problem10-full.pddl',
'problem10-half.pddl',
'problem11-full.pddl',
'problem11-half.pddl',
"problem04-full.pddl",
],
num_states=35000,
initial_sample_size=100,
distance_feature_max_complexity=5,
max_concept_grammar_iterations=3,
initial_concept_bound=10,
max_concept_bound=12,
concept_bound_step=2,
batch_refinement_size=10,
clean_workspace=False,
parameter_generator=add_domain_parameters,
feature_namer=feature_namer,
)
parameters = {
"problem02full": problem02full,
"problem03full": problem03full,
"problem03full_incremental": problem03full_incremental,
"guillem": guillem,
"visitall_incremental": visitall_incremental,
}.get(experiment_name or "test")
return generate_experiment(domain_dir, domain, **parameters)
