A linear programming approach to online set membership parameter estimation for linear regression models

This paper presents a new technique for online set membership parameter estimation of linear regression models affected by unknown-but-bounded noise. An orthotopic approximation of the set of feasible parameters is updated at each time step. The proposed technique relies on the solution of a suitable linear program, whenever a new measurement leads to a reduction of the approximating orthotope. The key idea for preventing the size of the linear programs from steadily increasing is to propagate only the binding constraints of these optimization problems. Numerical studies show that the new approach outperforms existing recursive set approximation techniques, while keeping the required computational burden within the same order of magnitude.