Precise piecewise affine models from input-output data

Formal design and analysis of embedded control software relies on mathematical models of dynamical systems, and such models can be hard to obtain. In this paper, we focus on automatic construction of piecewise affine models from input-output data. Given a set of examples, where each example consists of a d-dimensional real-valued input vector mapped to a real-valued output, we want to compute a set of affine functions that covers all the data points up to a specified degree of accuracy, along with a disjoint partitioning of the space of all inputs defined using a Boolean combination of affine inequalities with one region for each of the learnt functions. While traditional machine learning algorithms such as linear regression can be adapted to learn the set of affine functions, we develop new techniques based on automatic construction of interpolants to derive precise guards defining the desired partitioning corresponding to these functions. We report on a prototype tool, MOSAIC, implemented in Matlab. We evaluate its performance using some synthetic data, and compare it against known techniques using data-sets modeling electronic placement process in pick-and-place machines.