**17 commits**| Last commit

**≈ 21 months ago**

Ctrl K

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.

2

mentions

1

contributor

Get started

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_model
```

Available 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 env`

`source ./env/bin/activate`

`pip 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.

Keywords

Programming languages

License

- Open Access

</>Source code

Packages

Niki van Stein