Bespoke Learning to Generate Originally-Absent Training Data

This chapter motivates the need for attending to the class of problems that pertain to a mismatch between the proposed ambition of supervised learning of the relationship between a pair of variables, and the debilitating information deficiency that sometimes plagues such ambition—in the form of the lack of training data that is a requisite for the sought supervised learning. Multiple examples of such problems are presented, from across disciplines. The one and only solution that is then relevant, entails the learning of values of one variable—out of the relevant pair of variables—at design values of the other. Such learning is introduced as “bespoke learning”, and it is distinguished from supervised, unsupervised and reinforcement learning. Descriptions of the methodologies that accomplish the same are presented: in generic dynamical systems, to forecast future states by learning the evolution-driving function of the system; or to learn a parametrisation of the output variable that is a variably-long multivariate time series; in static systems, by formulating likelihood of the unknown system parameters given data, using accessible information. Summaries of the forthcoming empirical illustrations of such methodologies are included.