Tuning

Tuners

GEFEST provides api for 4 tuners from GOLEM:

Tuners comparation

	iopt	optuna	sequential	simultaneous
Nodes bypass	simultaneous	simultaneous	sequential	simultaneous
Backend	iOpt	Optuna	Hyperopt	Hyperopt
Multi objective	No	Yes	No	No

How to tune

To initialize tuner and run tuning, similarly to the optimisers, you need an OptimizationParams with defined TunerParams and also one or more Structure objects.

More details about OptimizationParams you can find in cases and API referece sections of this documentation.

Here we will take a closer look at several TunerParams attributes which may be unclear.

from gefest.tools.tuners.utils import percent_edge_variance

tuner_cfg = TunerParams(
    tuner_type='optuna',
    n_steps_tune=10,
    hyperopt_dist='uniform',
    variance_generator=percent_edge_variance,
    verbose=True,
    timeout_minutes=60,
)

• tuner_type is responsible for which tuner will be used.

• hyperopt_dist is the type of distribution from which random values will be taken during tuning. Available values are names of hpyeropt hp module finctions. Note: in GOLEM 0.4.0 part of tuners does not converts bounds into mu and sigma. Use ‘uniform’ to avoid invalid intervals.

• variance_generator is function that generates bounds of intervals from which random values should pe picked for all components of all point in structure. If normal distribution set they will be automatically converted into means and varicances.

percent_edge_variance function setes variance to spicific percent of average edge length for each polygon. This solution can be much more “greedy” than necessary, which can lead to many invalid intermediate variants during tuning. To improve fitness in fewer tuning steps, it is worth creating variance generation functions for selecting smaller intervals based on the conditions of a specific task.

Now that the OptimizationParams have been defined and some structures have been created, we can run tuning with couple lines of code:

from gefest.core.configs.optimization_params import OptimizationParams
from gefest.core.geometry.datastructs.structure import Structure
from gefest.tools.tuners.tuner import GolemTuner

structs: list[Structure] | Structure = ...
opt_params: OptimizationParams = ...

tuner = GolemTuner(opt_params)
tuned_structs = tuner.tune(structs)