Two Best Paper Awards for CWI at GECCO — CWI Amsterdam
Four researchers from the CWI Life Sciences and Health (LSH) group – Alexander Chebykin, Dazhuang Liu, Marco Virgolin and Peter AN Bosman (CWI/TU Delft) – as well as Tanja Alderliesten (LUMC) received the Best Paper Award in two titles from GECCO 2022. The awards were won in the titles: Neuroevolution for the paper “Evolutionary neural cascade research on superlattices“and the track Genetic Programming for Paper”Scalability Degeneracy in Multi-Objective Genetic Programming for Symbolic Regression“.
Evolutionary research in neural cascade through superlattices (Alexander Chebykin, Tanja Alderliesten, Peter AN Bosman)
Neural networks are powerful function approximators that have proven useful in a variety of fields. But there is always a desire to make them even more effective and efficient. One way to do this is to create waterfalls of different patterns. The researchers proposed a new scalable algorithm to create such cascades. This algorithm is efficient and can work with hundreds of models from any source: for example, pre-trained or created automatically for the target task by a neural architecture search algorithm. The resulting trade-off fronts of the cascades improve both the individual models and the cascades found by previous approaches.
Scalability Degeneracy in Multi-Objective Genetic Programming for Symbolic Regression (Dazhuang Liu, Marco Virgolin, Tanja Alderliesten, Peter AN Bosman)
Besides high predictive accuracy, interpretability can be a critical aspect for using machine learning in high-stakes applications (e.g. cancer treatment prediction). Genetic programming is a primary method for discovering accurate and interpretable ML models in the form of small symbolic expressions. Empirically, small models tend to be less accurate than large ones, i.e. there is a trade-off between accuracy and interpretability. Therefore, most researchers use GP in a multi-objective manner to simultaneously discover multiple models with different trade-offs. However, when used naively, MO-GP can “get stuck” with small patterns and fail to discover more accurate ones. Our researchers have found the root of this problem, which they have named “scalability degeneracy”. Then they devised a simple but perfect remedy. This resulted in a new algorithm, evoNSGA-IIwhich has proven to outperform previous MO-GP algorithms.