EN FR
EN FR
Overall Objectives
New Software and Platforms
Bilateral Contracts and Grants with Industry
Bibliography
Overall Objectives
New Software and Platforms
Bilateral Contracts and Grants with Industry
Bibliography


Section: New Results

Fitness landscape analysis for multi-objective optimization

Participants: F. Daolio, A. Liefooghe (external collaborators: Sébastien Verel, Univ. Littoral Côte d’Opale, France; Hernan Aguirre and Kiyoshi Tanaka, Shinshu Univ., Japan)

Computationally hard multi-objective combinatorial optimization problems are common in practice, and numerous evolutionary multi-objective optimization (EMO) algorithms have been proposed to tackle them. Our aim is to understand which (and how) problem features impact the search performance of such approaches. In [38] , we adopt a statistical approach, based on simple and multiple linear regression analysis, to enquire the expected running time of global SEMO with restart for identifying a (1+ε)−approximation of the Pareto set for small-size enumerable instances. Our analysis provides further insights on the EMO search behavior and on the most important features that characterize the difficulty of an instance for this class of problems and algorithms. In [31] , we consider two prototypical dominance-based algorithms: a global EMO strategy using an ergodic variation operator (GSEMO) and a neighborhood-based local search heuristic (PLS). Their respective runtime is estimated on a benchmark of combinatorial problems with tunable ruggedness, objective space dimension, and objective correlation (ρMNK-landscapes). In other words, benchmark parameters define classes of instances with increasing empirical problem hardness; we enumerate and characterize the search space of small instances. Our study departs from simple performance comparison to systematically analyze the correlations between runtime and problem features, contrasting their association with search performance within and across instance classes, for both chosen algorithms. A mixed-model approach then allows us to further generalize from the experimental design, supporting a sound assessment of the joint impact of instance features on EMO search performance. Next, in [28] , we analyse the behavior and compares the performance of MOEA/D, IBEA using the binary additive epsilon and the hypervolume difference indicators, and AεSεH as representative algorithms of decomposition, indicators, and ε-dominance based approaches for many-objective optimization. We use small MNK-landscapes to trace the dynamics of the algorithms generating high-resolution approximations of the Pareto optimal set. Also, we use large MNK-landscapes to analyze their scalability to larger search spaces. At last, in [39] , we report an experimental analysis on stochastic local search for approximating the Pareto set of bi-objective unconstrained binary quadratic programming problems. First, we investigate two scalarizing strategies that iteratively identify a high-quality solution for a sequence of sub-problems. Each sub-problem is based on a static or adaptive definition of weighted-sum aggregation coefficients, and is addressed by means of a state-of-the-art single-objective tabu search procedure. Next, we design a Pareto local search that iteratively improves a set of solutions based on a neighborhood structure and on the Pareto dominance relation. At last, we hybridize both classes of algorithms by combining a scalarizing and a Pareto local search in a sequential way. A comprehensive experimental analysis reveals the high performance of the proposed approaches, which substantially improve upon previous best-known solutions. Moreover, the obtained results show the superiority of the hybrid algorithm over non-hybrid ones in terms of solution quality, while requiring a competitive computational cost. In addition, a number of structural properties of the problem instances allow us to explain the main difficulties that the different classes of local search algorithms have to face.