Professor Ferrante Neri
Professor of Machine Learning and Artificial Intelligence, Head of the Nature Inspired Computing and Engineering (NICE) Research Group)
I am an all-round academic equally passionate about teaching and research. My teaching expertise is about Mathematical subjects for Computer Science while my research expertise lies at the intersection of Optimisation, Explainable AI, and Machine Learning. I am also happy to serve the community I work for, by undertaking managerial duties.
Before joining the faculty at Surrey I was at the University of Nottingham and previously I was at De Montfort University and at the University of Jyväskylä (Finland). I am a scholar in Optimisation and Computational Modelling. I work at the boundary between mathematical theory and practical engineering solutions. I’m the author of Linear Algebra for Computational Sciences and Engineering. I am also a very active editor for multiple journals including Information Sciences, Memetic Computing, and Integrated Computer-Aided Engineering.
In evolutionary biology, the term "fitness landscape" refers to the correlation between genotypes and reproduction success. In computer science, this concept is abstracted to general multi-variate functions in the context of optimisation and refers to the changes in the objective function values between neighbouring solutions. Intuitively, for a function in two variables, the 3D visualisation of the objective function displays "valleys", "peaks", "ridges", "plateaus", etc. thus identifying a "landscape". In multi-variate problems, since we cannot visualise the fitness landscape we must perform tests to analyse the features of the landscape and then understand/imagine it.
Fitness landscape analysis is important in Explainable AI since it produces precious pieces of information to design a successful optimiser. In the past decade, multiple studies about tests and metrics to analyse fitness landscapes have been proposed. However, how these metrics can be used to design a successful optimiser is still a mostly unaddressed topic. This talk presents some recent results about the design of Pattern Search algorithms based on fitness landscape analysis. The proposed analysis deals with the epistasis of the problem and searches for preferential search directions within the search space and presents three algorithmic implementations informed by this analysis. The relative advantages and disadvantages of each approach are highlighted and their results against traditional and modern optimisers are presented.