|April 11, 2022||Workshop paper submission deadline|
|April 25, 2022||Notification of acceptance|
|May 2, 2022||Camera-ready deadline|
|May 2, 2022||Author registration deadline|
|Speaker: Nelishia Pillay
Title: Transfer Learning in Automated Design Using Generation Hyper-Heuristics
Abstract: Transfer learning has proven to be effective in machine learning approaches such as deep neural networks. More recently, transfer learning has been used to improve the performance and scalability of evolutionary algorithms. The benefits of transfer learning in optimization include improvement in performance, improvement in runtimes, reduced data requirements and improved convergence. This talk will examine the role that transfer learning can play in generation hyper-heuristics for automated design in terms of both optimality and generality, highlighting its benefits in this area. Design decisions that generation hyper-heuristics are predominately used to automate include the creation of operators, generation of processes in algorithms and evolution of overall algorithms. The talk will present case studies for the relevant design decisions, illustrating how transfer learning can be used and the effectiveness of transfer learning for these design decisions. The talk will conclude by discussing future research directions.
Short Bio: Nelishia Pillay is a Professor at the University of Pretoria, South Africa. She holds the Multichoice Joint-Chair in Machine Learning and SARChI Chair in Artificial Intelligence for Sustainable Development. She is chair of the IEEE Technical Committee on Intelligent Systems Applications, Vice Chair of the IEEE Technical Committee on Evolutionary Computation, chair of the IEEE Task Force on Automated Algorithm Design, Configuration and Selection and chair of the IEEE CIS WCI subcommittee. She is associate editor for IEEE Computational Intelligence Magazine, IEEE Transactions on Emerging Topics in Computational Intelligence, Swarm and Evolutionary Computation and the Journal of Scheduling. Her research areas include hyper-heuristics, genetic programming and the automated design of machine learning and optimization techniques for sustainable development. These are the focus areas of the NICOG (Nature-Inspired Computing Optimization) research group which she has established.
The main objective of this workshop is to discuss hyper-heuristics and algorithm configuration methods for the automated generation and improvement of algorithms, with the goal of producing solutions (algorithms) that are applicable to multiple instances of a problem domain. The areas of application of these methods include optimization, data mining and machine learning. [1-18,23].
Automatically generating and improving algorithms by means of other algorithms has been the goal of several research fields, including Artificial Intelligence in the early 1950s, Genetic Programming since the early 1990s, and more recently automated algorithm configuration  and hyper-heuristics . The term hyper-heuristics generally describes meta-heuristics applied to a space of algorithms. While Genetic Programming has most famously been used to this end, other evolutionary algorithms and meta-heuristics have successfully been used to automatically design novel (components of) algorithms. Automated algorithm configuration grew from the necessity of tuning the parameter settings of meta-heuristics and it has produced several powerful (hyper-heuristic) methods capable of designing new algorithms by either selecting components from a flexible algorithmic framework [3,4] or recombining them following a grammar description .
Although most evolutionary algorithms are designed to generate specific solutions to a given instance of a problem, one of the defining goals of hyper-heuristics is to produce solutions that solve more generic problems. For instance, while there are many examples of evolutionary algorithms for evolving classification models in data mining and machine learning, a genetic programming hyper-heuristic has been employed to create a generic classification algorithm which in turn generates a specific classification model for any given classification dataset, in any given application domain . In other words, the hyper-heuristic is operating at a higher level of abstraction compared to how most search methodologies are currently employed; i.e., it is searching the space of algorithms as opposed to directly searching in the problem solution space , raising the level of generality of the solutions produced by the hyper-heuristic evolutionary algorithm. In contrast to standard Genetic Programming, which attempts to build programs from scratch from a typically small set of atomic functions, generative hyper-heuristic methods specify an appropriate set of primitives (e.g., algorithmic components) and allow evolution to combine them in novel ways as appropriate for the targeted problem class. While this allows searches in constrained search spaces based on problem knowledge, it does not in any way limit the generality of this approach as the primitive set can be selected to be Turing-complete. Typically, however, the initial algorithmic primitive set is composed of primitive components of existing high-performing algorithms for the problems being targeted; this more targeted approach very significantly reduces the initial search space, resulting in a practical approach rather than a mere theoretical curiosity. Iterative refining of the primitives allows for gradual and directed enlarging of the search space until convergence.
As meta-heuristics are themselves a type of algorithm, they too can be automatically designed employing hyper-heuristics. For instance, in 2007, Genetic Programming was used to evolve mate selection in evolutionary algorithms ; in 2011, Linear Genetic Programming was used to evolve crossover operators ; more recently, Genetic Programming was used to evolve complete black-box search algorithms [13,14,16], SAT solvers , and FuzzyART category functions . Moreover, hyper-heuristics may be applied before deploying an algorithm (offline)  or while problems are being solved (online) , or even continuously learn by solving new problems (life-long) . Offline and life-long hyper-heuristics are particularly useful for real-world problem solving where one can afford a large amount of a priori computational time to subsequently solve many problem instances drawn from a specified problem domain, thus amortizing the a priori computational time over repeated problem solving. Recently, the design of Multi-Objective Evolutionary Algorithm components was automated .
Very little is known yet about the foundations of hyper-heuristics, such as the impact of the meta-heuristic exploring algorithm space on the performance of the thus automatically designed algorithm. An initial study compared the performance of algorithms generated by hyper-heuristics powered by five major types of Genetic Programming . Another avenue for research is investigating the potential performance improvements obtained through the use of asynchronous parallel evolution to exploit the typical large variation in fitness evaluation times when executing hyper-heuristics .
Workshop papers must be submitted using the GECCO submission site. After login, the authors need to make sure that their role is set to "Submitter", select "Make a New Submission", then select "Workshop Paper", and then in the workshop submission form, the authors must select the workshop they are submitting to, namely "Workshop Evolutionary Computation for the Automated Design of Algorithms". To see a sample of the "Workshop Paper" submission form, go to the GECCO submission site, select "Make a New Submission", then select "Workshop Paper", and then select "view sample form".
Submitted papers must not exceed 8 pages (excluding references) and are required to be in compliance with the GECCO 2022 Paper Submission Instructions. It is recommended to use the same templates as the papers submitted to the main tracks. Each paper submitted to this workshop will be rigorously reviewed in a double-blind review process. In other words, authors should not know who the reviewers of their work are and reviewers should not know who the authors are. To this end, the following information is very important: Submitted papers should be ANONYMIZED. This means that they should NOT contain any element that may reveal the identity of their authors. This includes author names, affiliations, and acknowledgments. Moreover, any references to any of the author's own work should be made as if the work belonged to someone else. All accepted papers will be presented at the ECADA workshop and appear in the GECCO 2022 Conference Companion Proceedings. By submitting a paper, the author(s) agree that, if their paper is accepted, they will:
E-mail: email@example.comDr. López-Ibáñez is Senior Distinguished Researcher at the University of Málaga (Spain) and a Senior Lecturer (Associate Professor) in the Decision and Cognitive Sciences Research Centre at the Alliance Manchester Business School, University of Manchester, UK. He received the M.S. degree in computer science from the University of Granada, Granada, Spain, in 2004, and the Ph.D. degree from Edinburgh Napier University, U.K., in 2009. He has published 27 journal papers, 9 book chapters and 48 papers in peer-reviewed proceedings of international conferences on diverse areas such as evolutionary algorithms, ant colony optimization, multi-objective optimization, pump scheduling and various combinatorial optimization problems. His current research interests are experimental analysis and automatic design of stochastic optimization algorithms, for single and multi-objective optimization. He is the lead developer and current maintainer of the irace software package (http://iridia.ulb.ac.be/irace).
Daniel R. Tauritz is an Associate Professor in the Department of Computer Science and Software Engineering at Auburn University (AU), Interim Director and Chief Cyber AI Strategist of the Auburn Cyber Research Center, the founding Head of AU's Biomimetic Artificial Intelligence Research Group (BioAI Group), a cyber consultant for Sandia National Laboratories, a Guest Scientist at Los Alamos National Laboratory (LANL), and founding academic director of the LANL/AU Cyber Security Sciences Institute (CSSI). He received his Ph.D. in 2002 from Leiden University. His research interests include the design of generative hyper-heuristics and self-configuring evolutionary algorithms and the application of computational intelligence techniques in cyber security, critical infrastructure protection, and program understanding. He was granted a US patent for an artificially intelligent rule-based system to assist teams in becoming more effective by improving the communication process between team members.
E-mail: firstname.lastname@example.orgJohn R. Woodward is a Senior Lecturer, and head of The Operational Research Group at the Queen Mary University of London. Formerly he was a lecturer at the University of Stirling, within the CHORDS group and was employed on the DAASE project. Before that he was a lecturer for four years at the University of Nottingham. He holds a BSc in Theoretical Physics, an MSc in Cognitive Science and a PhD in Computer Science, all from the University of Birmingham. His research interests include Automated Software Engineering, particularly Search Based Software Engineering, Artificial Intelligence/Machine Learning and in particular Genetic Programming. He has over 50 publications in Computer Science, Operations Research and Engineering which include both theoretical and empirical contributions, and given over 50 talks at international conferences and as an invited speaker at universities. He has worked in industrial, military, educational and academic settings, and been employed by EDS, CERN and RAF and three UK Universities.