The Moving Target algorithm from scratch
A tutorial on the Moving Targets algorithm for constrained Machine Learning
This tutorial presents the key ideas in the Moving Targets algorithm, a decomposition based approach that relies on an external constraint solver to inject constraints and symbolic information in supervised learning.
Due to the use of a decomposition, the approach can deal with any supervised learning method and (in principle) any kind of constraint, though in practice one needs expertise both in choosing the correct solution technique and the correct constraint formulation. Overall, the main attractiveness of the approach is that of providing an alternative avenue for introducing symbolic information in ML.
The tutorial is focused on explaining the main ideas in the algorithm in a visual fashion: as a consequence, only a toy example of learning with constraints will be considered here.
Familiarity with Supervised Learning and with some method for constrained optimization