Much research in evolutionary computation focuses on the design and application of new algorithm variants for obtaining better results for hard application problems. For evolution strategies, a class of evolutionary algorithms, I will briefly discuss a few variants and their performance, and then illustrate a typical application example from the domain of industrial engineering design (such as ships, cars, etc.).
I will argue, however, that the traditional, expert-driven algorithm development approach has its limitations, and that there are suitable approaches towards an automatic configuration of evolutionary algorithms in general, and evolution strategies in particular. This is motivated by some early work on the (hyper-parameter ) optimization of genetic algorithms. Recent research in my group resulted in the configurable CMA-evolution strategy, which makes 4.308 variants of evolution strategies possible - most of which have never been tried before. This generalization can not only be used for searching through configuration space, but the optimization results can also be analyzed using data-driven approaches, resulting in a better understanding of the search mechanism.
In my perspective, this is only the beginning, since the results clearly indicate that the algorithm configuration should also be able to switch during the optimization run, and that the switching strategy could add an additional benefit for the algorithm’s convergence runtime. Ultimately, it would be great if the algorithm could continuously learn the changing requirements for module activations, which requires internal machine learning mechanisms such as multi-armed bandit or reinforcement learning approaches. For both topics, I will discuss our first results and outline the potential for further research. The talk concludes by providing a short overview of research topics and industrial projects in my research group.
Thomas Bäck is professor of Computer Science at the Leiden Institute of Advanced Computer Science, Leiden University, Netherlands, since 2002. He received his PhD in Computer Science from Dortmund University, Germany, in 1994, and was leader of the Center for Applied Systems Analysis at the Informatik Centrum Dortmund until 2000. Until 2007, Thomas was also CTO of NuTech Solutions, Inc. (Charlotte, NC), where he gained ample experience in solving real-world problems in optimization and data analytics, by working with global enterprises in the automotive and other industry sectors.
Thomas received the IEEE Computational Intelligence Society (CIS) Evolutionary Computation Pioneer Award for his contributions in synthesizing evolutionary computation (2015), was elected as a fellow of the International Society of Genetic and Evolutionary Computation (ISGEC) for fundamental contributions to the field (2003), and received the best dissertation award from the "Gesellschaft für Informatik" in 1995.
Thomas has more than 300 publications, as well as two books on evolutionary algorithms: Evolutionary Algorithms in Theory and Practice (1996), Contemporary Evolution Strategies (2013). He is co-editor of the Handbook of Evolutionary Computation and the Handbook of Natural Computing. Thomas is also Co-Editor-in-Chief of the Natural Computing Series (Springer) and the journal Theoretical Computer Science C (Elsevier), and editorial board member of various journals.
Thomas’ research interests are in foundations and applications of evolutionary computation, efficient global optimization, and multiple objective optimization. His recent research also addresses optimization methods in machine learning, e.g., for hyperparameter optimization, as well as supervised and unsupervised machine learning for smart industry applications.