Data Driven Evolutionary Optimization of Complex Systems: Big Data Versus Small Data

This invited talk Existing evolutionary algorithms typically assume that there are explicit objective functions available for fitness evaluations. In the real world, such explicit objective functions may not be available in many cases. For example, many industrial optimization problems such as structural design [8] need to perform computationally very intensive numerical simulations, such as computational fluid dynamic simulations or finite element analysis, where a large set of partial differential equations must be solved. In many process industry optimization problems, no explicit models exist for describing the relationship between the final quality of the product and the decision variables, such as temperature and humidity. Thus, only historical experimental data can be used for optimization. There are also cases where only factual data can be collected. A good example of such optimization problems is trauma systems design [11], where only patient records are available for optimization. For solving such optimization problems, evolutionary optimization can be conducted only using a data-driven approach. The main challenges in data-driven evolutionary optimization can roughly be divided into two categories according to the amount of available data, namely, small data and big data. The lack of data can mainly be attributed to the fact that data acquisition is very expensive, either computationally or costly. In [11], data-driven evolutionary optimization problems are divided into two paradigms, one termed off-line data-driven optimization, where no new data can be actively sampled, and the other on-line data-driven optimization, where a small number of new data points can