PPSN 2020 Workshop on (Multimodal) Multi-Objective Optimization


(Multimodal) Multi-Objective Optimization (MMMOO):

Challenges, Characteristics, and Peculiarities


-- to be held hybrid (onsite/online) at PPSN 2020 --



 


Free Access to the Workshop


The workshop will be made publicly available (for free). If you would like to attend the workshop, contact us via e-mail (e.g., kerschke@uni-muenster.de) and we will send you the link.


 


Schedule


All talks are scheduled for 10 minutes plus 10 minutes Q&A.


  • Session 1 (Sunday, September 6, 11:00am - 12:30pm):
    • 11:00 -- 11:10 Workshop organizers (onsite):
      • Welcome to Session 1
    • 11:10 -- 11:30 Marcus Gallagher (University of Queensland, Australia) (online):
      • On the Challenges of Multimodal Single-Objective Optimization
    • 11:30 -- 11:50 Hisao Ishibuchi (Southern University of Science and Technology, China) (online):
      • Multi-Modal Multi-Objective Optimization
      • Abstract: In this short talk, first the popularity of multi-modal multi-objective optimization research is analyzed using the Web of Science data base in comparison with the popularity of many-objective optimization research. It is clearly shown that the number of multi-modal multi-objective papers increased drastically last year. Next, multi-modal multi-objective problems are visually explained using some experimental results. Diversification mechanisms in the decision space are usually combined with multi-objective algorithms for multi-modal multi-objective optimization. Using experimental results, two negative effects of such a diversity increase attempt are explained: One is the slowdown of the convergence speed in the objective space, and the other is the decrease of the diversity in the objective space. Finally, some future research directions are explained, which include discussions on the diversity decrease in the objective space, handing of the convergence-diversity balance in both the objective and decision spaces, discussions on performance indicators in the decision space, and proposal of new test problems.
    • 11:50 -- 12:10 Estefania Yap (University of Melbourne, Australia) (online):
      • Instance Spaces in Continuous Multi-Objective Optimisation

        (Co-authors: Mario A. Muñoz, Kate Smith-Miles)
      • Abstract: Evolutionary multi-objective optimization algorithms have proven to be popular for solving multi-objective optimization problems, with ongoing research and development in this space leading to an abundance of new algorithms.Furthermore, there is a lack of understanding into the relationship between algorithm performance and problem instances. As such, identifying a suitable algorithm for a given problem instance continues to be a difficult task. In this preliminary work, we utilize instance space analysis for multi-objective continuous problems for the first time. Previously, visualizations in the discrete multi-objective space were generated, using newly identified features. We extend these features into the continuous space, generate new multi-objective features, and combine them with single-objective features to generate new instance spaces. We explore the validity of translating discrete features into the continuous space, and identify relationships between features of benchmark problem suites and algorithm performance.
    • 12:10 -- 12:30 Lennart Schäpermeier (University of Münster, Germany) (onsite):
      • A Dashboard for Visualizing Multi-Objective Landscapes: The Perspective Matters

  • Session 2 (Sunday, September 6, 1:30pm - 3:00pm):
    • 1:30 -- 1:35 Workshop organizers (onsite):
      • Welcome to Session 2
    • 1:35 -- 1:55 Gabriela Ochoa (University of Stirling, UK) (online):
      • Complex Networks in Search and Optimisation
    • 1:55 -- 2:15 Stef Maree (University of Amsterdam / CWI, Netherlands) (onsite):
      • Uncrowded Hypervolume-Based Multi-Objective Optimization

        (Co-authors: Timo M. Deist, Tanja Alderliesten, Peter Bosman)
      • Abstract: Domination-based multi-objective (MO) evolutionary algorithms (EAs), such as the NSGA-II, are very effective algorithms to quickly obtain a good set of solutions. However, these methods tend to loose selection pressure when the entire population is non-dominated, which then causes stagnation in terms of convergence to Pareto-optimal solutions. Hypervolume-based MO optimization can overcome this fundamental limitation of domination-based MO EAs. In this talk, I will give an overview of (our) recent work on this topic, which is based on the uncrowded hypervolume indicator. I will address how MO problems can be efficiently solved with a single-objective EA, while showing convergence to Pareto optimality. I will furthermore discuss how this approach can be used to perform gradient-based MO optimization, and show how we can use this approach to make the selection of a desired solution from the approximation set simpler and more insightful for the decision maker.
    • 2:15 -- 3:00 Panel Discussion

 


Scope of the Workshop


In practical applications, multi-objective (MO) optimization is usually treated secondarily due to its rather deterrent complexity/difficulty (compared to single-objective (SO) optimization problems). Therefore, practitioners in general scalarize their problems, e.g., by optimizing weighted sums of the underlying objectives. A major reason for such behavior is the much lesser tangibility of MO problems; it is extremely challenging to imagine interaction effects between > 2 decision variables and > 2 objectives simultaneously (let alone visualize them within a single plot).


Even researchers usually limit themselves to visualizing only the Pareto fronts of MO problems, i.e., the image of the set of MO global optima. As a result, our "knowledge" about MO problems is highly influenced by our understanding of SO problems. For instance, it is well-known that multimodality can be very challenging in SO optimization. Thus, for a long time, research simply inferred that such structures cause similar problems for optimizers in the MO setting. In consequence, such structures have regularly been considered for the design of MO benchmark problems. Yet, recent works have shown that multimodality might in fact even facilitate MO optimization.


 


Topics of Interest


In an attempt to reduce our knowledge deficit in this particular domain, the proposed workshop shall provide a platform for researchers to actively exchange ideas that improve our understanding of (multimodal) MO continuous optimization problems. We therefore welcome contributions related to the following non-exclusive list of topics:


  • Characteristics of continuous and combinatorial MO optimization problems.
  • High-level landscape characteristics (such as ridges, plateaus, etc.) as well as exploratory landscape features.
  • Empirical and theoretical results on the transferability of structural properties from SO to MO problems.
  • Multiobjectivization strategies for SO problems.
  • Techniques for visualizing landscapes of (multimodal) MO problems.
  • Algorithms and/or algorithm building blocks (e.g., operators, selection mechanisms) that are capable of handling or exploiting discovered challenges in problem structures (such as multimodality, ill-conditioned landscapes, etc.).
  • Consequences for the design of benchmark problems and evaluation of existing test problem suites.

 


Workshop Organizers


 


Disclaimer Regarding Covid-19


This workshop is part of the PPSN 2020 in Leiden, The Netherlands. Due to the restrictions resulting from the Covid-19 pandemic, it will be held as a mixture of onsite and online talks.


 


Further Notes


Please help us raising awareness for this workshop and distribute the link to this website (http://www.erc.is/go/mmmoo2020).