vsr-behavior-characterization

In this small project we apply an Unsupervised Learning technique (KMeans) to find recurrent behaviors in a dataset of Voxel-Based Soft Robots (VSR).

Structure of the project

The folder src contains the core of the project, in particular the file behavior_classification.py is responsible for the execution of the experiments. The other files are auxiliary functions/classes used to clean the code.

The folder dataset contains the data which we arw willing to clusterize. best.[0-9].txt contain the predictors on which we apply KMeans, please refer to this webpage for an overview of their meaning. The files supervised_clusters*.py contain two different versions of verification datasets that we prepared by hand via the manual inspection of videos available alongside the datasets.

Example run and output

Navigate to the root folder of the project (you should see both src and dataset). To run an experiment, use the following command

python3 src/behavior_classification.py N

where N is the desired number of clusters. Note that at the moment we support only N=3 and N=4 since these are the only values for which we managed to generate two supervised datasets "by hand", in which we classified the behavior of the robots according to our opinion. This allows us to select the best combination of the weights we introduced in our approach (and which we wish to optimize in order to obtain improved results). Therefore, new supervised datasets are needed to increase the domain of N. Have a look at the files dataset/supervised_clusters*.py to see how we represented the supervised clusters.

The script tries a wide set of possible values for the weights, and for each combination produces the following output:

This first step may last up to one hour, depending on your device. When all the combination have been tested, an ouput similar to following is printed (we set N=3 for this example):

As you can see you receive immediately a clear overview of the results for the optimal combination, along with an in-depth comparison between supervised and unsupervised clusters (i.e. clusters built "by hand" and clusters built by the algorithm KMeans).

Authors

• Francesco Andreuzzi
• Luca Filippi

Implementation details

We coded the experiment using Python, along with the scientific libraries NumPy and scikit-learn.

References

1. Ferigo et al. 2021, Beyond body shape and brain: evolving the sensory apparatus of voxel-based soft robots.
2. Hastie, Tibshirani, Friedman, 2009, An introduction to statistical learning.
3. Medvet et al. 2020, Design, validation, and case studies of 2d-vsr-sim, an optimization-friendly simulator of 2-d voxel-based soft robots.
4. Medvet et al. 2021, Biodiversity in evolved voxel-based soft robots.
5. Panday et al. 2018, Feature weighting as a tool for unsupervised feature selection.