New Evolutionary Computing Developments Optimize Complex Problem Solving

The platform is based on evolutionary algorithms that optimize the search for solutions to complex scientific and engineering problems. These results are applicable to many fields, like molecular chemistry, materials resistance, robotics or games theory.

This research line was developed under the leadership of professors José María Peña Sánchez and Antonio Latorre de la Fuente, and has resulted in several publications and final-year projects, including Manuel Zaforas Martín's report on the implementation of this platform, known as Multiple Offspring Sampling (MOS).

Evolutionary algorithms are a family of algorithms within the artificial intelligence (AI) world that are useful for solving non-linear and very complex search and optimization problems, where there is a trade-off between the quality of the solutions and the required computation time. The members of this class of problems range from classical mathematical challenges to real-world scientific and engineering problems. Generally, we can use these methods to tackle problems about which little is known a priori and that would otherwise be intractable.

These methods are inspired by the theory of evolution postulated by Darwin in 1859. Continuing the biological metaphor, a population of candidate solutions "evolves" as if they were individuals until they arrive at the best possible solution. Several techniques implement these principles in one way or another, and each technique has its particularities, strengths and weaknesses.

New methodology: MOS

What the UPM School of Computing researchers have managed to do is develop a methodology, called MOS, that can simultaneously use and intelligently combine different evolutionary techniques and get the best out of each one. This way, MOS is capable of operating with several evolutionary models, such as the popular genetic algorithms (GA), estimation of distribution algorithms (EDA) based on probabilistic models or more recent techniques like differential evolution (DE).

The MOS methodology abstracts the each evolutionary algorithm's mechanisms for generating new offspring and encapsulates them in what is called a technique. Additionally, it defines new concepts like technique "quality". Quality is determined by how good or bad the behaviour of the technique is depending on any characteristic that is to be improved in the new individuals. Also quality dynamically defines each technique's participation in the evolutionary process. This way, the techniques work together and achieve better results than they would have individually, solving complex problems faster and more accurately.

This research has been possible thanks to a partnership with CeSViMa, also based at the UPM's Montegancedo Campus. CeSViMa owns Spain's second-most powerful computer. Thanks to experiments run on the machine that has a massive computational capacity, the researchers were able to analyse the strengths of these algorithms. They were put to the task of solving very complex mathematical problems that no conventional computer would be able to solve.

At present the line of research is still open with several doctoral theses in the making. The group is working on both adding new techniques and improving the technique hybridization processes.