Section: Research Program

Decomposition-based Optimization

Given the large scale of the targeted optimization problems in terms of the number of variables and objectives, their decomposition into simplified and loosely coupled or independent subproblems is essential to raise the challenge of scalability. The first line of research is to investigate the decomposition approach in the two spaces and their combination, as well as their implementation on ultra-scale architectures. The motivation of the decomposition is twofold: first, the decomposition allows the parallel resolution of the resulting subproblems on ultra-scale architectures. Here also several issues will be addressed: the definition of the subproblems, their coding to allow their efficient communication and storage (checkpointing), their assignment to processing cores, etc. Second, decomposition is necessary for solving large problems that cannot be solved (efficiently) using traditional algorithms. Indeed, for instance with the popular NSGA-II algorithm the number of non-dominated solutions (A solution x dominates another solution y if x is better than y for all objectives and there exists at least one objective for which x is strictly better than y.) increases drastically with the number of objectives leading to a very slow convergence to the Pareto front (A Pareto Front is the set of non-dominated solutions.). Therefore, decomposition-based techniques are gaining a growing interest. The objective of Bonus is to investigate various decomposition schema and cooperation protocols between the subproblems resulting from the decomposition to generate efficiently global solutions of good quality. Several challenges have to be addressed: how to define the subproblems (decomposition strategy), how to solve them to generate local solutions (local rules), and how to combine these latter with those generated by other subproblems and how to generate global solutions (cooperation mechanism), and how to combine decomposition strategies in more than one space (hybridization strategy)? These challenges, which are in the line with the CIS Task Force (IEEE CIS Task Force, created in 2017 on Decomposition-based Techniques in Evolutionary Computation.) on decomposition will be addressed in the decision as well as in the objective space.

The decomposition in the decision space can be performed following different ways according to the problem at hand. Two major categories of decomposition techniques can be distinguished: the first one consists in breaking down the high-dimensional decision vector into low-dimensional and easier-to-optimize blocks of variables. The major issue is how to define the subproblems (blocks of variables) and their cooperation protocol: randomly vs. using some learning (e.g. separability analysis), statically vs. adaptively, etc. The decomposition in the decision space can also be guided by the type of variables i.e. discrete vs. continuous. The discrete and continuous parts are optimized separately using cooperative hybrid algorithms  [47]. The major issue of this kind of decomposition is the presence of categorial variables in the discrete part [33]. The Bonus team is addressing this issue, rarely investigated in the literature, within the context of vehicle aerospace engineering design. The second category consists in the decomposition according to the ranges of the decision variables. For continuous problems, the idea consists in iteratively subdividing the search (e.g. design) space into subspaces (hyper-rectangles, intervals, etc.) and select those that are most likely to produce the lowest objective function value. Existing approaches meet increasing difficulty with an increasing number of variables and are often applied to low-dimensional problems. We are investigating this scalability challenge (e.g. [10]). For discrete problems, the major challenge is to find a coding (mapping) of the search space to a decomposable entity. We have proposed an interval-based coding of the permutation space for solving big permutation problems. The approach opens perspectives we are investigating [7], in terms of ultra-scale parallelization, application to multi-permutation problems and hybridization with metaheuristics.

The decomposition in the objective space consists in breaking down an original MaOP into a set of cooperative single-objective subproblems (SOPs). The decomposition strategy requires the careful definition of a scalarizing (aggregation) function and its weighting vectors (each of them corresponds to a separate SOP) to guide the search process towards the best regions. Several scalarizing functions have been proposed in the literature including weighted sum, weighted Tchebycheff, vector angle distance scaling, etc. These functions are widely used but they have their limitations. For instance, using weighted Tchebycheff might do harm diversity maintenance and weighted sum is inefficient when it comes to deal with nonconvex Pareto Fronts  [40]. Defining a scalarizing function well-suited to the MaOP at hand is therefore a difficult and still an open question being investigated in Bonus  [6], [5]. Studying/defining various functions and in-depth analyzing them to better understand the differences between them is required. Regarding the weighting vectors that determine the search direction, their efficient setting is also a key and open issue. They dramatically affect in particular the diversity performance. Their setting rises several issues mainly: how to determine their number according to the available computational ressources? when (statically or adaptively) and how to determine their values? Weight adaptation is one of our main concerns that we are addressing especially from a distributed perspective. They correspond to the main scientific objectives targeted by our bilateral ANR-RGC BigMO project with City University (Hong Kong). The other challenges pointed out in the beginning of this section concern the way to solve locally the SOPs resulting from the decomposition of a MaOP and the mechanism used for their cooperation to generate global solutions. To deal with these challenges, our approach is to design the decomposition strategy and cooperation mechanism keeping in mind the parallel and/or distributed solving of the SOPs. Indeed, we favor the local neighborhood-based mating selection and replacement to minimize the network communication cost while allowing an effective resolution [5]. The major issues here are how to define the neighborhood of a subproblem and how to cooperatively update the best-known solution of each subproblem and its neighbors.

To sum up, the objective of the Bonus team is to come up with scalable decomposition-based approaches in the decision and objective spaces. In the decision space, a particular focus will be put on high dimensionality and mixed-continuous variables which have received little interest in the literature. We will particularly continue to investigate at larger scales using ultra-scale computing the interval-based (discrete) and fractal-based (continuous) approaches. We will also deal with the rarely addressed challenge of mixed-continuous including categorial variables (collaboration with ONERA). In the objective space, we will investigate parallel ultra-scale decomposition-based many-objective optimization with ML-based adaptive building of scalarizing functions. A particular focus will be put on the state-of-the-art MOEA/D algorithm. This challenge is rarely addressed in the literature which motivated the collaboration with the designer of MOEA/D (bilateral ANR-RGC BigMO project with City University, Hong Kong). Finally, the joint decision-objective decomposition, which is still in its infancy  [49], is another challenge of major interest.