Optimisation problems are part of everyday life: from planning efficient supply chains to designing complex production processes and even creating timetables for schools and universities. Modern Artificial Intelligence and Machine Learning also rely on solving optimisation problems in order to minimise training errors on large datasets.
News
New Approaches in Optimisation Research
Mathematical Analysis Tool for Complex Problems
A research team led by Prof. (FH) Dr Rubén Ruiz Torrubiano, Head of the Engineering Responsible AI Systems degree programme at IMC University of Applied Sciences Krems, together with Prof. (FH) Dipl. Ing. Dr techn. Deepak Dhungana, Prof. (FH) Dipl.-Ing. Dr techn. Sarita Paudel and Assoc. Prof. (FH) Himanshu Buckchash, PhD, has developed an innovative approach to systematically analyse and select so-called metaheuristics – intelligent approximation methods for particularly complex optimisation problems. The results of their work have been published in the renowned journal Algorithms (MDPI).
Metaheuristics are inspired by phenomena in nature and technology, such as evolution, the cooling of materials, or the behaviour of animal colonies. They make it possible to solve highly complex problems with great accuracy, which could not be tackled efficiently using classical methods, as these would take far too long to arrive at practical solutions. However, a key challenge lies in the sheer number of available metaheuristics, making the choice of the right method a practical problem in itself. Which algorithm is best suited for which problem?
Concrete Guidance for Choosing the Right Algorithm
In their new paper “Modelling Local Search Metaheuristics using Markov Decision Processes”, the research team proposes modelling metaheuristics as stochastic agents within a Markov Decision Process. This framework not only enables the theoretical analysis of the convergence behaviour of individual algorithms, but also provides concrete guidance as to which metaheuristic is best suited for a given problem. The method can be applied flexibly to various algorithms such as genetic algorithms – a particular focus for future research.