IEEE WCCI2026-CEC
Workshop on Generative and Adaptive
Surrogate Models for Evolutionary Algorithms

Abstract:

This keynote presents our recent research developments on surrogate-assisted evolutionary optimization for computationally expensive combinatorial optimization problems. The talk focuses on hybrid optimization frameworks developed in our research projects, where machine learning-based surrogate models are integrated within evolutionary algorithms to reduce the computational burden associated with expensive fitness evaluations during the search process.

The presentation first discusses our work on edge-based facility location problems, where objective-function evaluations are computationally demanding due to repeated edge-decomposition procedures. Surrogate models are employed in this context to approximate expensive fitness evaluations and accelerate the evolutionary search.

The talk then presents our recent studies on electric vehicle charging scheduling problems, in both offline and online settings, and unrelated parallel machine scheduling problems with maintenance constraints. These problems are addressed through hierarchical optimization frameworks, where evolutionary algorithms optimize upper-level decision variables, while lower-level subproblems are solved using mathematical optimization or pseudo-polynomial dynamic programming techniques. Since evaluating candidate solutions becomes computationally demanding throughout the evolutionary search, surrogate models are incorporated to estimate objective-function values and improve computational efficiency.

The keynote also presents our recent methodological investigations on improving surrogate-model performance through advanced feature-engineering mechanisms and problem-aware learning representations designed to better capture structural characteristics of optimization instances.

Short Bio:

Lhassane Idoumghar is a full Professor (Exceptional-class 2) at the University of Haute-Alsace in France. Since 01/2020, he is Director of IRIMAS Institute (Institut de recherche en Informatique, Mathématique, Automatique et Signal). 01/2018 - 12/2019, he was elected as Deputy Director of IRIMAS Institute, and Director of the IRIMAS - Computer Science Research department, University of Haute-Alsace, Mulhouse. He is Head of the OMEGA (Optimization by MEtaheuristics, alGorithms and modelization) research team. Until 2017, he was Associate Director of LMIA Laboratory (Laboratoire de Mathématiques, Informatique et Application), University of Haute Alsace, Mulhouse. In 2015, he was the co-founder of the UHA 4.0 school which is a new innovative school in terms of pedagogical. Since, he is Projects Director at UHA 4.0 school, University of Haute-Alsace, Mulhouse. In 2012, he obtained his accreditation to supervise research from University of Haute-Alsace. In 2003, He obtained his PhD at the University of Henri Poincaré - Nancy 1, France. In 2000-2003 he worked in TDF-C2R Broadcasting and wireless Research Center. His research activities include Artificial Intelligence, Evolutionary Algorithm, massively parallel and distributed metaheuristics, Optimization and uncertain optimization by hybrid metaheuristics. In addition to his academic and professional pursuits, he earned his Private Pilot License (PPL) at the Trois Frontières Aero-Club (AC3F) in Habsheim, France.

Abstract:

TBA

Abstract:

What if a surrogate model had already read every algorithm ever written? Large language models, trained on vast corpora of scientific code and mathematical reasoning, bring an unprecedented prior over the space of optimization algorithms, one that requires no explicit fitting, generalizes across problem classes, and can be queried not through an analytic acquisition function, but through natural language reasoning.

In this talk, we explore how this perspective reframes LLMs as reasoning-aware surrogate models within evolutionary computation. We present LLaMEA, a framework in which LLMs act as both generator and adaptive search operator, evolving optimization algorithms through iterative code generation and performance feedback. We discuss how the same paradigm extends to Bayesian optimization algorithm design via LLaMEA-BO, where LLMs autonomously discover novel BO routines that generalize across benchmark suites — rivalling and in some cases surpassing hand-crafted state-of-the-art methods.

Beyond benchmarks, we show how Exploratory Landscape Analysis can serve as a bridge between real-world problem structure and LLM-guided algorithm generation, enabling informed, problem-aware optimization in settings where evaluation is expensive and prior knowledge is scarce.

The talk raises as many questions as it answers: When does an LLM's world-prior help, and when does it mislead? How do we benchmark systems that generate their own solvers? And what does it mean to explain the decisions of a surrogate that reasons in code?

Short Bio:

Niki van Stein is Associate Professor of Explainable AI at the Leiden Institute of Advanced Computer Science (LIACS), Leiden University, where she leads research at the intersection of explainable AI, automated algorithm design and metaheuristic optimization. Her work spans surrogate-assisted optimization, AutoML, benchmarking methodology, and more recently the use of large language models to automatically generate and adapt algorithms. She is the lead developer of LLaMEA, an LLM-driven evolutionary framework for algorithm discovery that received the Silver Humies Award at GECCO 2025 and won several code competitions. She has co-organized international seminars and workshops on LLMs for evolutionary computation, is chair of the ECTA, AGENTICS and EXPLAINS conferences and serves in the editorial board of ECJ. Her research finds application in engineering design, predictive maintenance, and scientific optimization.