Visualization in Multiobjective Optimization
Tutorial by Bogdan Filipic and Tea Tušar
IEEE Congress on Evolutionary Computation (CEC 2017)
Donostia - San Sebastián, 5 June 2017
Multiobjective optimization algorithms usually produce a set of trade-off solutions approximating the Pareto front where no solution from the set is better than any other in all objectives (this is called an approximation set). While there exist many measures to assess the quality of approximation sets, no measure is as effective as visualization, especially if the Pareto front is known and can be visualized as well. Visualization in evolutionary multiobjective optimization is relevant in many aspects, such as estimating the location, range, and shape of the Pareto front, assessing conflicts and trade-offs between objectives, selecting preferred solutions, monitoring the progress or convergence of an optimization run, and assessing the relative performance of different algorithms. This tutorial will provide a comprehensive overview of methods used in multiobjective optimization for visualizing either individual approximation sets or the probabilistic distribution of multiple approximation sets through the empirical attainment function (EAF).
The tutorial will build on the well-attended 2016 tutorial edition and extend it with a taxonomy of visualization methods and ten additional visualization methods not presented last year: aggregation trees, distance-based and dominance-based mappings, level diagrams with asymmetric norm, moGrams, polar plots, tetrahedron coordinates model, trade-off region maps, treemaps, visualization following Shneiderman mantra, and 3D-RadVis. We will demonstrate how each of the introduced methods visualizes the benchmark approximation sets. In addition, animations will be used where possible to show the additional capabilities of the visualization methods. Tutorial attendees will become aware of the many ways of visualizing approximation sets and the EAF, and learn about the advantages and limitations of each method.
Parallel and distributed evolutionary algorithms
Tutorial by El-Ghazali Talbi
IEEE Congress on Evolutionary Computation (CEC 2017)
Donostia - San Sebastián, 5 June 2017
Parallel and distributed computing can be used in the design and implementation of evolutionary algorithms for speedup the search, improve the quality of the obtained solutions, improve the robustness of the obtained solutions, and solve large scale problems.
From the algorithmic design point of view, we will present the main parallel models for evolutionary algorithms (algorithmic level, iteration level, solution level). We will address also:
- Parallel hybrid models with exact methods.
- Parallel models for multi-objective optimization.
- Illustrations solving large challenging applications in networks, logistics and transportation and bioinformatics.
From the implementation point of view, we here concentrate on the parallelization of evolutionary algorithms on general-purpose parallel and distributed architectures, since this is the most widespread computational platform. The rapid evolution of technology in terms of processors (GPUs, multi-core), networks (Infiniband), and architectures (GRIDs, clusters, Clouds) make those architectures very popular nowadays.
Different architectural criteria which affect the efficiency of the implementation will be considered: shared memory / distributed memory, homogeneous / heterogeneous, dedicated / non dedicated, local network / large network. Indeed, those criteria have a strong impact on the deployment technique employed such as load balancing and fault-tolerance.
Finally, some software frameworks for parallel evolutionary algorithms such as PARADISEO are presented. Those frameworks allow the design of parallel and hybrid metaheuristics for mono-objective and multi-objective optimization, and the transparent implementation on different parallel and distributed architectures using adapted middleware.
Meta-Model Assisted (Evolutionary) Optimization
Tutorial by Boris Naujoks, Jörg Stork, Martin Zaefferer, and Thomas Bartz-Beielstein
14th International Conference on Parallel Problem Solving from Nature (PPSN 2016)
Edinburgh, 18 September 2016
Meta-model assisted optimization is a well-recognized research area. When the evaluation of an objective function is expensive, meta-model assisted optimization yields huge improvements in optimization time or cost in a large number of different scenarios. Hence, it is extremely useful for numerous real-world applications. These include, but are not limited to, the optimization of designs like airfoils or ship propulsion systems, chemical processes, biogas plants, composite structures, and electromagnetic circuit design.
This tutorial is largely focused on evolutionary optimization assisted by meta-models, and has the following aims: Firstly, we will provide a detailed understanding of the established concepts and distinguished methods in meta-model assisted optimization. Therefore, we will present an overview of current research and open issues in this field. Moreover, we aim for a practical approach. The tutorial should enable the participants to apply up-to-date meta-modelling approaches to actual problems at hand. Afterwards, we will discuss typical problems and their solutions with the participants. Finally, the tutorial offers new perspectives by taking a look into areas where links to meta-modelling concepts have been established more recently, e.g., the application of meta-models in multi-objective optimization or in combinatorial search spaces.
Please download the presentation here.
Visualization in Multiobjective Optimization
Tutorial by Bogdan Filipic and Tea Tušar
The Genetic and Evolutionary Computation Conference (GECCO 2016)
Denver, 20 July 2016
Multiobjective optimization algorithms usually produce a set of trade-off solutions approximating the Pareto front where no solution from the set is better than any other in all objectives (this is called an approximation set). While there exist many measures to assess the quality of approximation sets, no measure is as effective as visualization, especially if the Pareto front is known and can be visualized as well. This tutorial provides a comprehensive overview of methods used in multiobjective optimization for visualizing either individual approximation sets or the probabilistic distribution of multiple approximation sets through the empirical attainment function (EAF).
This tutorial is among the first comprehensive presentations of visualization methods for multiobjective optimization and is valuable to students, researchers, algorithm designers and end users dealing with multiobjective optimization problems.
Contents of the tutorial:
- Requirements for visualization methods
- Benchmark approximation sets that can be used for comparing visualization methods in a similar way as performance metrics and benchmark test problems are used for comparing optimization algorithms
- Two groups of methods used for visualizing individual approximation sets (for each method we will demonstrate its outcome on the benchmark approximation sets and discuss its advantages and limitations):
- General methods (not designed for multiobjective optimization): scatter plot matrix, bubble chart, radial coordinate visualization, parallel coordinates, heatmaps, Sammon mapping, neuroscale, self-organizing maps, principal component analysis and isomap
- Specific methods (able to handle the unique features of approximation sets): distance and distribution charts, interactive decision maps, hyper-space diagonal counting, two-stage mapping, level diagrams, hyper-radial
visualization, Pareto shells, seriated heatmaps, multidimensional scaling and prosections
- Visualization of EAF values and differences
We demonstrate how each of the introduced methods visualizes the benchmark approximation sets. In addition, for some of the methods (such as parallel coordinates, prosections and direct volume rendering) animations are used to show their additional capabilities.
Please download the presentation here.