2 Best Paper Awards for CWI at GECCO

4 researchers of CWI’s Life Sciences and Well being (LSH) group – Alexander Chebykin, Dazhuang Liu, Marco Virgolin, and Peter A.N. Bosman (CWI/TU Delft) – along with Tanja Alderliesten (LUMC) obtained the Best Paper Award in two tracks of GECCO 2022. The awards have been received within the tracks: Neuroevolution for the paper (*2*) and the observe Genetic programming for the paper “Evolvability Degeneration in Multi-Objective Genetic Programming for Symbolic Regression“.

Abstracts

Evolutionary Neural Cascade Search throughout Supernetworks (Alexander Chebykin, Tanja Alderliesten, Peter A. N. Bosman)

Neural networks are highly effective perform approximators which have confirmed helpful in a wide range of domains. However there’s all the time a want to make them much more efficient and environment friendly. One option to obtain that is by creating cascades of various fashions. The researchers got here up with a novel evolutionary algorithm for creating such cascades. This algorithm is environment friendly and may work with a whole bunch of fashions from any supply: e.g. pre-trained, or created for the goal activity robotically by a Neural Structure Search algorithm. The ensuing trade-off fronts of cascades enhance each upon the person fashions and the cascades discovered by the earlier approaches.

Evolvability degeneration in multi-objective genetic programming for symbolic regression (Dazhuang Liu, Marco Virgolin, Tanja Alderliesten, Peter A.N. Bosman)

Moreover excessive predictive accuracy, interpretability might be a vital facet for the usage of machine studying in high-stakes purposes (e.g., most cancers remedy prediction). Genetic programming is a primary methodology to find correct and interpretable ML fashions within the type of small symbolic expressions. Empirically, smaller fashions are usually much less correct than bigger ones, i.e., there exists a trade-off between accuracy and interpretability. Consequently, most researchers use GP in a multi-objective style to concurrently uncover a number of fashions with completely different trade-offs. Nonetheless, when used naively, MO-GP can “get stuck” with small fashions, and fail to find extra correct ones. Our researchers have discovered the foundation of this drawback, which they named “evolvability degeneration”. Subsequent, they designed a easy however spot-on treatment. This resulted in a brand new algorithm, evoNSGA-II, which was discovered to outperform earlier MO-GP algorithms.