Invited Keynote Talk of Thomas Bäck at Genetic Programming Theory & Practice XX, June 1-3

On June 2, Thomas Bäck will give an invited keynote talk at the 20th Genetic Programming Theory & Practice (GPTP) conference, taking place at Michigan State University, East Lansing, MI. It is a small, invitation-only workshop hosted 2023 by the BEACON Center for the Study of Evolution in Action.  

The idea behind Genetic Programming is to use evolutionary algorithm principles to evolve successively better performing computer programs using the concepts of a population of candidate programs and evolutionary operators such as selection, recombination, and mutation. For evaluating the quality of the evolving programs, an objective function needs to be defined.  

Thomas’ presentation is entitled “Automated Algorithm Configuration for Expensive Optimization Tasks” and addresses two key topics which seem, at first, unrelated to Genetic Programming: Expensive optimization problems, requiring significant computational resources for a single objective function evaluation, and methods that can be used for algorithm configuration and hyperparameter optimization. In the presentation, he shows how this topic and the related algorithms relate to Genetic Programming, by giving three examples:  

The first example directly applies a hyperparameter optimization method called mixed-integer parallel efficient global optimization (MIP-EGO) to optimize the hyperparameters of Grammatical Evolution, a paradigm strongly related to Genetic Programming. The results illustrate improvements in performance of Grammatical Evolution of up to 168% for a typical test problem.  

The second example shows a first approach towards using Genetic Programming to evolve test functions that have the same characteristics as real-world optimization problems, but can be evaluated much faster and therefore serve as candidates for creating and training optimization algorithms that can then solve the real-world problem very efficiently. 

The third example is focusing on the potential of Genetic Programming to generate explainable AI models for high-dimensional data. As shown on the example of high-dimensional sensor time series data from aircraft engine operations, Genetic Programming can generate compact, human-readable models that explain the relation between critical sensors and the exhaust gas temperature of the engine.