Optimizers

The table below displays all the available optimizers and the correlation between them and a parameter-type / multi-objective function:

Optimizer	Parameter type			Multi-objective function
	Continuous	Discrete	Categorical	Chimera [1]	Weighted Sum	Pareto Front Optimization
Falcon GPBO*
Semopt*
EDBO+
Grid
RandomSearch

* For Enterprise users only

** Initial proposals are random. Click on the specific optimizer for more details.

Falcon GPBO

Type: Bayesian optimizer

Description: Gaussian Process priors are probably the most standard choice of surrogate model for problems that are conceivably optimized with a relatively small number of observations. However, its computational cost--it scales cubically in the number of observations--hinders its usage for more challenging problems.

References: [1]

Grid - 0.1.0

Type: (Quasi)-Random and DOE strategy

Description: A method for optimizing parameters by doing exhaustive search of the parameter space in a grid-like manner. Powerful method for exploring the whole space, with the possibility to use continuous, discrete and categorical parameters.

References: [1]

(03.2020 - in-house implementation)

RandomSearch - 0.1.0

Type: (Quasi)-Random and DOE strategy

Description: Method that uses random sampling to optimize the hyperparameter space. It can be used with continuous, discrete or categorical parameters.

(03.2020 - in-house implementation)

Semopt

Type: Transfer Learning

SeMOpt is Atinary’s transfer-learning algorithm to accelerate optimization processes by leveraging existing knowledge learned from past experiments and existing data. It automatically identifies key information and guides the optimization algorithms to the most promising parameter region from the start.

References: [1]

EDBO+

Type: Bayesian optimizer

Custom Bayesian optimizer developed by Dr. Abigail G. Doyle s group, built upon BoTorch.
This optimizer uses a Gaussian process surrogate model and hypervolume improvement for its acquisition function.
Currently supports up to 500000 combinations of parameter options.

References: [1]

Get Optimizer specifications from SDK

Here is a sample code to get the current optimizers specifications:

import scientia_sdk
import pprint
pp = pprint.PrettyPrinter(indent=4)

API_KEY = # Add your API-KEY here
ENDPOINT_URL = # Add relevant host, cf. "API Endpoints" section

configuration = scientia_sdk.Configuration(
    host = ENDPOINT_URL,
    api_key = {
        'api_key': API_KEY,
    }
)
configuration.access_token = None  # temporary fix when using the SDK

with scientia_sdk.ApiClient(configuration) as api_client:
    api_optimizer = scientia_sdk.OptimizerApi(api_client)
    api_response = api_optimizer.optimizers_specification()
    pp.pprint([i["function"] for i in api_response.to_dict()["objects"]])