DoE2Vec is a self-supervised approach to learn exploratory landscape analysis features from design of experiments. The model can be used for downstream meta-learning tasks such as learninig which optimizer works best on a given optimization landscape. Or to classify optimization landscapes in function groups.
The approach uses randomly generated functions and can also be used to find a "cheap" reference function given a DOE. The model uses Sobol sequences as the default sampling method. A custom sampling method can also be used. Both the samples and the landscape should be scaled between 0 and 1.
`pip install doe2vec`Afterwards you can use the package via:
from doe2vec import doe_modelAvailable models can be viewed here: https://huggingface.co/BasStein
A model name is build up like BasStein/doe2vec-d2-m8-ls16-VAE-kl0.001
Where d is the number of dimensions, 8 the number (2^8) of samples, 16 the latent size, VAE the model type (variational autoencoder) and 0.001 the KL loss weight.
Example code of loading a huggingface model
obj = doe_model(
2,
8,
n= 50000,
latent_dim=16,
kl_weight=0.001,
use_mlflow=False,
model_type="VAE"
)
obj.load_from_huggingface()
#test the model
obj.plot_label_clusters_bbob()python3.8 -m venv envsource ./env/bin/activatepip install -r requirements.txt
To generate the artificial function dataset for a given dimensionality and sample size run the following code
from doe2vec inport doe_model
obj = doe_model(d, m, n=50000, latent_dim=latent_dim)
if not obj.load():
obj.generateData()
obj.compile()
obj.fit(100)
obj.save()Where d is the number of dimensions, m the number of samples (2^m) per DOE, n the number of functions generated and latent_dim the size of the output encoding vector.
Once a data set and encoder has been trained it can be loaded with the load() function.