"""Functions to create, run and visualize optimization benchmarks.
TO-DO:
- Add other benchmark sets:
- finish medium scale problems from https://arxiv.org/pdf/1710.11005.pdf, Page 34.
- add scalar problems from https://github.com/AxelThevenot
- Add option for deterministic noise or wiggle.
"""
import numpy as np
from estimagic import batch_evaluators
from estimagic.algorithms import AVAILABLE_ALGORITHMS
from estimagic.optimization.optimize import minimize
from pybaum import tree_just_flatten
from estimagic.parameters.tree_registry import get_registry
[docs]def run_benchmark(
problems,
optimize_options,
*,
batch_evaluator="joblib",
n_cores=1,
error_handling="continue",
max_criterion_evaluations=1_000,
disable_convergence=True,
):
"""Run problems with different optimize options.
Args:
problems (dict): Nested dictionary with benchmark problems of the structure:
{"name": {"inputs": {...}, "solution": {...}, "info": {...}}}
where "inputs" are keyword arguments for ``minimize`` such as the criterion
function and start parameters. "solution" contains the entries "params" and
"value" and "info" might contain information about the test problem.
optimize_options (list or dict): Either a list of algorithms or a Nested
dictionary that maps a name for optimizer settings
(e.g. ``"lbfgsb_strict_criterion"``) to a dictionary of keyword arguments
for arguments for ``minimize`` (e.g. ``{"algorithm": "scipy_lbfgsb",
"algo_options": {"convergence.relative_criterion_tolerance": 1e-12}}``).
Alternatively, the values can just be an algorithm which is then benchmarked
at default settings.
batch_evaluator (str or callable): See :ref:`batch_evaluators`.
n_cores (int): Number of optimizations that is run in parallel. Note that in
addition to that an optimizer might parallelize.
error_handling (str): One of "raise", "continue".
max_criterion_evaluations (int): Shortcut to set the maximum number of
criterion evaluations instead of passing them in via algo options. In case
an optimizer does not support this stopping criterion, we also use this as
max iterations.
disable_convergence (bool): If True, we set extremely strict convergence
convergence criteria by default, such that most optimizers will exploit
their full computation budget set by max_criterion_evaluations.
Returns:
dict: Nested Dictionary with information on the benchmark run. The outer keys
are tuples where the first entry is the name of the problem and the second
the name of the optimize options. The values are dicts with the entries:
"params_history", "criterion_history", "time_history" and "solution".
"""
if isinstance(batch_evaluator, str):
batch_evaluator = getattr(
batch_evaluators, f"{batch_evaluator}_batch_evaluator"
)
opt_options = _process_optimize_options(
optimize_options,
max_evals=max_criterion_evaluations,
disable_convergence=disable_convergence,
)
minimize_arguments, keys = _get_optimization_arguments_and_keys(
problems, opt_options
)
raw_results = batch_evaluator(
func=minimize,
arguments=minimize_arguments,
n_cores=n_cores,
error_handling=error_handling,
unpack_symbol="**",
)
processing_arguments = []
for name, raw_result in zip(keys, raw_results):
processing_arguments.append(
{"optimize_result": raw_result, "problem": problems[name[0]]}
)
results = batch_evaluator(
func=_process_one_result,
arguments=processing_arguments,
n_cores=n_cores,
error_handling="raise",
unpack_symbol="**",
)
results = dict(zip(keys, results))
return results
def _process_optimize_options(raw_options, max_evals, disable_convergence):
if not isinstance(raw_options, dict):
dict_options = {}
for option in raw_options:
if isinstance(option, str):
dict_options[option] = option
else:
dict_options[option.__name__] = option
else:
dict_options = raw_options
default_algo_options = {}
if max_evals is not None:
default_algo_options["stopping.max_criterion_evaluations"] = max_evals
default_algo_options["stopping.max_iterations"] = max_evals
if disable_convergence:
default_algo_options["convergence.relative_criterion_tolerance"] = 1e-14
default_algo_options["convergence.relative_params_tolerance"] = 1e-14
default_algo_options["convergence.relative_gradient_tolerance"] = 1e-14
out_options = {}
for name, _option in dict_options.items():
if not isinstance(_option, dict):
option = {"algorithm": _option}
else:
option = _option.copy()
algo_options = {**default_algo_options, **option.get("algo_options", {})}
algo_options = {k.replace(".", "_"): v for k, v in algo_options.items()}
option["algo_options"] = algo_options
if isinstance(option.get("algo_options"), dict):
option["algo_options"] = {**default_algo_options, **option["algo_options"]}
else:
option["algo_options"] = default_algo_options
out_options[name] = option
return out_options
def _get_optimization_arguments_and_keys(problems, opt_options):
kwargs_list = []
names = []
for prob_name, problem in problems.items():
for option_name, options in opt_options.items():
algo = options["algorithm"]
if isinstance(algo, str):
if algo not in AVAILABLE_ALGORITHMS:
raise ValueError(f"Invalid algorithm: {algo}")
else:
valid_options = AVAILABLE_ALGORITHMS[algo]._algorithm_info.arguments
else:
valid_options = algo._algorithm_info.arguments
algo_options = options["algo_options"]
algo_options = {k: v for k, v in algo_options.items() if k in valid_options}
kwargs = {**options, **problem["inputs"]}
kwargs["algo_options"] = algo_options
kwargs_list.append(kwargs)
names.append((prob_name, option_name))
return kwargs_list, names
def _process_one_result(optimize_result, problem):
"""Process the result of one optimization run.
Args:
optimize_result (OptimizeResult): Result of one optimization run.
problem (dict): Problem specification.
Returns:
dict: Processed result.
"""
_registry = get_registry(extended=True)
_criterion = problem["noise_free_criterion"]
_start_x = problem["inputs"]["params"]
_start_crit_value = _criterion(_start_x)
if isinstance(_start_crit_value, np.ndarray):
_start_crit_value = (_start_crit_value**2).sum()
_is_noisy = problem["noisy"]
_solution_crit = problem["solution"]["value"]
# This will happen if the optimization raised an error
if isinstance(optimize_result, str):
params_history_flat = [tree_just_flatten(_start_x, registry=_registry)]
criterion_history = [_start_crit_value]
time_history = [np.inf]
batches_history = [0]
else:
history = optimize_result.history
params_history = history["params"]
params_history_flat = [
tree_just_flatten(p, registry=_registry) for p in params_history
]
if _is_noisy:
criterion_history = np.array([_criterion(p) for p in params_history])
if criterion_history.ndim == 2:
criterion_history = (criterion_history**2).sum(axis=1)
else:
criterion_history = history["criterion"]
criterion_history = np.clip(criterion_history, _solution_crit, np.inf)
batches_history = history["batches"]
time_history = history["runtime"]
return {
"params_history": params_history_flat,
"criterion_history": criterion_history,
"time_history": time_history,
"batches_history": batches_history,
"solution": optimize_result,
}