Research in optimization tends to focus on how to improve the design or configuration of their methods from a theoretical perspective or motivated by empirical studies of algorithm behavior on either specific individual problems or broader benchmarking problem suites. Our Dagstuhl Seminar takes a different perspective – we wish to improve the algorithm practitioner, arming them with a recipe book for improving any one of a broad class of optimization methods (black-box methods) on any given problem, particularly focusing on real-world problems.

In order to establish such a recipe book, we must look to real-world problems and first understand how they arrive on the desk of an optimization practitioner. The formulation of the problem itself is often (though not always) “messy” with many crucial details missing (such as what the objective(s) are, what the decision space is, and what the constraints are)! But even when such details are apparently in place, the practitioner must remember they are choices, and it will be well to understand who has made those choices and why. At a second level, one must understand how many times the optimization problem must be solved, with what frequency, on what machines, with what data and simulators, and so on. At another level, the practitioner must understand uncertainty in the problem, how the solution is represented, how it will be used in practice, what current solution methods are used, and how they were arrived at. By considering these and a large host of other factors, the practitioner can then understand how they might go about leveraging all that domain knowledge to tailor an optimization method for that problem to achieve the best possible tuned method. The challenge for us all is to be able to do this for any given problem and to have tailoring strategies that work well for any (or a wide range of) black-box methods. We know that such general tailoring methods exist in broad outline in the advice given in old textbooks [1], but it is a long time since they were updated for modern optimization methods and modern infrastructure, such as our modern methods of benchmarking algorithms.

The purpose of this Dagstuhl Seminar is to audit the current experience that exists in the field, pool it, and start to describe that experience in a way that goes beyond anecdote (although anecdotes can be useful) and moves us toward advice or recipes that actually work in practice. To achieve this, we aim to bring together researchers from a diverse range of backgrounds involved in real-world optimization from the customer side, the algorithm design or theoretical side, or practitioners who have already experienced successes and failures in the endeavor of tailoring algorithms to problems. The format of our seminar will be structured so that we hear from the cross-section of experienced and inexperienced practitioners, algorithm designers and customers, right from the start. We will then self-organize to achieve our aim of sharing our pooled experience and shaping it into practical, searchable and up-to-date advice on leveraging domain knowledge.

[1] L. Davis, Handbook of genetic algorithms. Chapman & Hall, London, 1991.

Creative Commons BY 4.0

Joshua D. Knowles, Katherine M. Malan, Elena Raponi, and Vanessa Volz