Big Multiobjective Optimization – BigMO

Following the evolution of modern computational science, the field of evolutionary optimization is shifting rapidly to the big era where the large-scale nature of applications implies big optimization models, with a large number of decision variables and conflicting objective functions. Big multi-objective optimization can be viewed following three inter-dependent axis: scalability, heterogeneity and uncertainty. These research axes give rise to difficult challenges being beyond the ability of commonly used optimization algorithms. In particular, we argue that standard techniques from evolutionary multi-objective optimization, although being arguably among the most efficient computing methodologies for dealing directly with multiple objectives, have to be deeply renovated in order to accelerate their widespread uptake in nowadays applications. The goal of this project is to foster the next generation of optimizers and to contribute to the intelligent design of advanced and effective optimization approaches attacking big multi-objective optimization problems and setting up an innovative and solid fundamental understanding of their characteristics. More specifically, our goal is to investigate the foundations of novel algorithms based on the concept of evolutionary decomposition, with the main challenge of designing autonomous mechanisms leveraging techniques at the crossroads of evolutionary computing and machine learning, and exploiting the tremendous computing power offered by modern parallel platforms in order to overcome the difficulty inherent to the problem dimension and complexity.

Accordingly, this project will adopt and set up the foundations of an evolutionary decomposition framework to deal with big multi-objective optimization problems. The corner stone of this framework is to systematically decompose a multi-objective problem into a number of single-objective or simple multi-objective sub-problems. Different search procedures with some loose inter-communications are employed for optimizing these sub-problems in a cooperative manner. Advanced tools from landscape analysis and parameter tuning will be fundamental in order to better understand the dynamics and the performance of the designed framework, and the impact of cooperation. Ultimately, we target the design of adaptive techniques, inspired by statistical learning and reinforcement learning, in order to adjust the behavior of each search procedure dynamically during execution. Such an approach will make the framework extremely effective to accommodate different existing optimizers and reduce the burden for users while being as much as possible generic to considered big optimization problems.

While such an evolutionary decomposition framework will be valuable to reduce the complexity of big multi-objective optimization, it is nonetheless necessary to deal with the high number of sub-problems and the cost that can be inherent to the evaluation of new candidate solutions generated concurrently for each sub-problem. To deal with this two issues, we propose to investigate two interdependent research paths. First, we rely on surrogate meta-models in order to predict the solutions quality without systematically computing their (expensive) objective value(s). In this respect, decomposition opens new research opportunities such as incorporating ensemble of surrogates to cope with different sub-problems characteristics. Second, the decentralized nature of decomposition makes it very appealing to use high-performance computing in order to distribute the underlying computations on the increasingly available compute facilities. While being a challenging issue due to the heterogenous and complex nature of modern large-scale compute platforms, incorporating parallelism is a strong and unique feature of the BigMO project as it is though to be a highly coupled with the design of the target decomposition framework and not in a separate manner.