Eight ESRs have been recruited. All of them stayed full-time until the last day of the project. They have completed all planned training activities, including training on the core algorithms and theories required to conduct research on Learning to Optimise, reading skills, presentation skills and networking. First, they attended a number of academic modules offered by the Universities, broadening their knowledge. Second, they were embedded into research training programmes at the Universities of Birmingham and Leiden and at the Honda Research Institute Europe and NEC Lab Europe, gaining first-hand research and industrial skills and experiences. Third, they attended a number of transferable skill training activities. Fourth, they have practiced their skills through conference presentations, posters, public talks, social media posts, etc.
The Learning to Optimise framework for solving complex engineering optimisation problems requires (a) advancing learning, (b) advancing optimisation, and (c) developing novel approaches to integrate learning with optimisation. The following progress has been achieved by the ECOLE project:
a) Learning:
• Learning from imbalanced data: Real world problems frequently suffer from data imbalances, where data of a given type is available in much more limited quantity than data of other types. Machine learning algorithms struggle to learn well from such data. We investigated the adequacy of different class imbalance techniques to deal with this problem and applied them to classify mesh quality in aerodynamics simulation of a benchmark vehicle body, which allows to realize a fast and resource efficient design process. Algorithms to automatically tune parameters of class imbalance techniques were proposed.
• Text representation: Many real-world problems involve making predictions based on text data. For that, existing machine learning algorithms use text representations that are not human-interpretable, leading to lack of human trust in the results. We proposed a novel human-interpretable representation for text and applied it to interpretable user profiling. We validated the method on user generated product review data and showed promising explanation about users’ preference for feature optimization of product design.
• Prediction approaches: We developed a novel prediction approach with demonstrated improvements in medical data analysis for patient status prediction. We integrated patient meta data, such as disease and demographic information, to model correlations of patient time series, by which prediction depends on not only historic patient information but also information of related patients.
b) Optimisation:
• Supporting automobile design: We collected data from various steps of vehicle development process to create a challenging test framework for new optimisation and learning methods. We improved a novel geometric representation based on cutting-edge deep learning technology for practical application on 3D (car) shapes. Such representation has shown to enable efficient optimisation of the design of a benchmark vehicle body.
• Optimisation under uncertainty: Real-world optimisation frequently has to deal with problems whose features are not entirely and accurately known beforehand. We developed a novel optimisation algorithm able to improve optimisation of such challenging problems.
c) Integration:
• Learning to optimise for dynamic problems: Real-world optimisation problems frequently suffer changes over time. We proposed a new algorithm to learn from the optimisation results obtained before a change to more efficiently optimise after the change.
• We also studied using multi-objective evolutionary optimisation algorithm for learning, where we can consider multiple objectives simultaneously without assigning weights to objectives.
