The problem-specific customization of meta-heuristic optimizers is an essential step for focusing the search and achieving scalability of standard of-the-shelf techniques to large-scale industrially relevant problems. Historically, this would have been achieved through the careful manual, design of various meta-heuristic components by an optimization expert. The automatic analysis of data (using machine learning approaches) offers an alternative approach and promises the opportunity to move beyond the capabilities of a human in identifying and capturing complex relationships, and integrating this information fully into the search process. Relevant data in an optimization setting may arise in a variety of forms including e.g. historical data capturing observed co-variation between pairs of decision variables, data describing previously seen distributions of individual variables, or collections of candidate solutions for simplified / different instances of the problem. In this talk, I will discuss these opportunities using a range of examples that span from structural biology over logistics to cluster analysis, specifically focusing on a range of open problems.
