Four researchers of CWI’s Life Sciences and Health (LSH) group – Alexander Chebykin, Dazhuang Liu, Marco Virgolin, and Peter A.N. Bosman (CWI/TU Delft) - together with Tanja Alderliesten (LUMC) received the Best Paper Award in two tracks of GECCO 2022. The awards were won in the tracks: Neuroevolution for the paper "Evolutionary Neural Cascade Search across Supernetworks" and the track Genetic programming for the paper "Evolvability Degeneration in Multi-Objective Genetic Programming for Symbolic Regression".
Evolutionary Neural Cascade Search across Supernetworks (Alexander Chebykin, Tanja Alderliesten, Peter A. N. Bosman)
Neural networks are powerful function approximators that have proven useful in a variety of domains. But there is always a desire to make them even more effective and efficient. One way to achieve this is by creating cascades of different models. The researchers came up with a novel evolutionary algorithm for creating such cascades. This algorithm is efficient and can work with hundreds of models from any source: e.g. pre-trained, or created for the target task automatically by a Neural Architecture Search algorithm. The resulting trade-off fronts of cascades improve both upon the individual models and the cascades found by the previous approaches.
Evolvability degeneration in multi-objective genetic programming for symbolic regression (Dazhuang Liu, Marco Virgolin, Tanja Alderliesten, Peter A.N. Bosman)
Besides high predictive accuracy, interpretability can be an essential aspect for the use of machine learning in high-stakes applications (e.g., cancer treatment prediction). Genetic programming is a prime method to discover accurate and interpretable ML models in the form of small symbolic expressions. Empirically, smaller models tend to be less accurate than larger ones, i.e., there exists a trade-off between accuracy and interpretability. Consequently, most researchers use GP in a multi-objective fashion to simultaneously discover multiple models with different trade-offs. However, when used naively, MO-GP can “get stuck” with small models, and fail to discover more accurate ones. Our researchers have found the root of this problem, which they named “evolvability degeneration”. Next, they designed a simple but spot-on remedy. This resulted in a new algorithm, evoNSGA-II, which was found to outperform previous MO-GP algorithms.