Stefan Menzel, Yew Ong, Yaochu Jin, Bernhard Sendhoff, "Special Session on Generative AI and Heuristic Optimization", IEEE Congress on Evolutionary Computation, 2023.

Generative AI and Large Language Models as groundbreaking technological innovations promise to redefine the boundaries of contemporary science and engineering. Their capabilities to produce e.g., context-sensitive text, knowledge-based answers, software code, images, music and 3D assets from text prompts and image inputs create opportunities for a manifold of disciplines. Through the training on large data, these models conserve knowledge, identify hidden patterns and even reason over other data modalities. In the context of evolutionary computation, generative models are of interest as they allow to map different input types, e.g., text prompts or images, to variations of output classes, e.g., text, images, 3D objects, which then can be utilized in evolutionary optimization tasks, such as prompt engineering, image preference optimization or design optimization. Especially the capabilities of evolutionary optimizers to cope with noise are advantageous to balance the novelty of generated solutions to realistic boundary conditions. From another perspective, Large Language Models are also explored as evolutionary optimization methods incorporating selection and mutation operators for single as well as multi-/many-objective optimization. Since generative models are data driven, i.e., trained on large data sets, they are also ideal candidates for transfer learning and reuse of history data, and thus providing solutions of good quality in similar design spaces, however may limit at the same time the exploration capability of the design space. Another highly relevant field is data synthesis, i.e., generating missing data samples in the problem space. Here, evolutionary optimization can help to provide a guided search under given computational budget limits.