ML-Methods for Identifying Features of Glob. Optimization

Evolutionary algorithms are nature-inspired learning, search and optimization methods that take the natural evolutionary process of a species as a model to adapt in the best possible way to environmental conditions. Methodical developments in combination with constantly increasing computing resources lead to the fact that more and more complex and higher-dimensional tasks can be solved with evolutionary algorithms. However, almost all of the currently considered optimization and machine learning tasks are stationary which means that the optimization or modeling goal does not change during an algorithm run. This project deals with different methodological approaches to move into the non-stationary domain. Although so far applied almost exclusively to stationary tasks, evolutionary approaches are an ideal starting point, since natural evolution itself is highly non-stationary. A species that loses its adaptability to new environmental conditions by adapting too greedily to currently prevailing conditions would become extinct just as well at this would happen if environmental conditions change too rapidly for a species to adapt. A sufficiently high presence of genetic diversity is therefore just as important in nature in terms of adaptability as it is for non-stationary optimization and modeling. Involving biological expertise, the Austrian project group will research implicit methods such as self- adaptive process extensions that maintain adaptability through a constant support of new genetic diversity without having to abandon what has already been learned. Age-layered population structures represent a promising methodological starting point here. So far, approaches of this kind have been mainly used to reduce premature convergence by a constant supply of new genetic diversity via the young age layers. In the course of the present project, these approaches will be extended in the direction of permanent adaptation to changing goals and framework conditions, and concepts will be explored as to how and which strategic method information can be transferred from the older to the younger age layers. The cooperating group from the Siberian State University of Science and Technology will go in the same direction with explicit approaches that change algorithm parameters or switch between algorithms at runtime based on fitness indicators.

This project explored how optimization algorithms can learn to adapt to changing conditions over time, much like how evolution helps species survive in shifting environments. Traditional optimization methods assume that a problem, such as planning delivery routes or scheduling production, stays fixed until solved. In reality, conditions change constantly: new jobs arrive, routes take longer than expected, or machines break down. The project focused on how evolutionary algorithms, which imitate natural selection and survival of the fittest, can be extended to handle such dynamic situations. A central theme was how to prevent a kind of dynamic collapse that can occur when an optimizer becomes too narrow or greedy for short-term gains. The team explored several ways to build resilience. Diversity-preserving measures like layered population structures let an optimizer maintain a steady mix of new and old solutions and avoid overfocusing. For problems where solutions vary in complexity (such as symbolic regression models), continuous pruning kept mathematical expressions compact and interpretable. And in practical tests like dynamic warehouse stacking, the researchers showed that short-sighted strategies could cause the system to grind to a halt. This highlighted the need to balance short-term efficiency with long-term stability and to include warning systems that alert human operators to trends the optimizer cannot avoid. As the project progressed, attention turned to the interaction between optimization algorithms and the machine learning models used to estimate solution quality. When an algorithm repeatedly avoids certain operating conditions, the predictive models may learn that these conditions are poor choices, even if they are not. This creates a feedback loop that can reinforce bias and limit performance. The team developed methods to detect, monitor, and mitigate such effects, ensuring that learning-based evaluations remain accurate and balanced over time. These findings also pointed to a broader challenge: even when evaluations are correct, optimization often involves trade-offs between competing objectives. Finally, the project addressed these trade-offs directly. Improving one measure, such as cost or speed, often worsens another, such as quality or sustainability. The researchers developed mathematical tools for comparing and predicting the progress of algorithms under such conflicting demands, including refined acquisition functions (selection criteria that weigh predicted qualities against uncertainties) and probability models that estimate the chances of further improvement. In summary, the project combined principles from evolutionary optimization, mathematical modelling, and artificial intelligence to improve how optimization systems respond to change. The developed methods help algorithms analyse the dynamics of problems and adapt their behaviour accordingly.