2018 - 2021
University of Lille / Inria
Machine Learning and Decomposition Techniques for Large-scale Multi-objective Optimization
The thesis is in the context of the bilateral France / Hong Kong BigMO project (big multi- objective optimization), funded by the French National Research Agency (ANR), in cooperation with City University of Hong Kong. The BigMO project aims at fostering the next generation multi-objective algorithms. The research challenges deal with the large-scale, expensive, and heterogeneous nature of big optimization problems, and the increasing availability of parallel computing facilities. The research conducted in this thesis project will contribute in setting up the foundations of a decomposition framework dealing with large-scale black-box optimization problems with heterogeneous objective functions. The corner stone of this framework is to systematically decompose a multi-objective problem into a number of smaller sub-problems both in the decision space and in the objective space. The first challenge is to accurately define the sub-problems and their inter- dependencies. Different search procedures are then to be designed for optimizing these sub-problems in a cooperative manner. Techniques from machine learning (e.g., surrogate meta-models, multi-armed bandits), landscape analysis and online/offline automatic algorithm selection and configuration are to be considered. Additionally, the decentralized nature of decomposition makes it very appealing to use distributed computing in order to parallelize the underlying computations. Depending on the advances, the research will focus on the following topics:
- Autonomous decomposition-based optimization integrating landscape analysis, feature- based algorithm selection, online/offline algorithm configuration.
- Benchmarking decomposition-based and machine learning-enhanced algorithms on cross-domain large-scale multi-objective optimization problems.
- Large-scale decomposition and cooperation in the decision space and in the objective space.
- Machine learning for boosting and generalizing decomposition-based optimization.
- High-level (algorithmic level, distributed model) and low-level (heterogeneity, parallel implementation and real-testbeds) parallelism.