public.loading

10th Nov 26

CISUC Connect | Distinguished Talk with Wolfgang Banzhaf

Speaker: Wolfgang Banzhaf

Date: 20th of November, 2026

Time: 10:00am

Place: Room B1

Presentation title: “Phenotypic Entropy and Simplicity Bias - Towards an Entropic Law for Evolving AI Systems” 

Short bio:

Wolfgang Banzhaf holds the first endowed chair dedicated to Evolutionary Computation in the United States, the John R. Koza Chair for Genetic Programming, in the Department of Computer Science and Engineering at Michigan State University. Previously, he was a University Research Professor in the Department of Computer Science, Memorial University of Newfoundland, where he also served as head of department from 2003 to 2009 and from 2012 to 2016. After a PhD and postdoc in Physics, he worked as a research scientist at Mitsubishi Electric in Japan and the US before joining academia in 1993 as associate professor of Applied Computer Science at the Technical University of Dortmund in Germany. His research interests are in the field of bio-inspired computing, notably evolutionary computing and complex adaptive systems. Author of a number of technical books and founder of the scholarly journal “Genetic Programming and Evolvable Machines” he teaches computational foundations of artificial intelligence and bio-inspired computing at Michigan State University.

Abstract:

Genetic programming has a well-known tendency to find simple solutions, sometimes too simple, as when a symbolic regression task results in a very simple equation. We explain this tendency using the structure of the genotype–phenotype map rather than by fitness evaluation. To this end we employ a linear GP model system for Boolean functions, where neutral networks can be counted exactly. This leads us to define a phenotypic entropy akin to Boltzmann’s entropy based on
the number of microstates W of a thermodynamic system. We show that the evolutionary process via neutral drift pulls a population toward high-entropy, simple phenotypes. Selection can hold a population against this entropic pull, but once selection relaxes, the population drifts back. We argue that this entropic simplicity bias is a statistical law of evolving systems, with consequences for parsimony pressure in GP and possible parallels in neural networks and large language models.

This website uses cookies to improve your experience. Read More