Estimates of the approximation error using Rademacher complexity: Learning vector-valued functions (Q1008456)

The authors improve a previously known estimate of the approximation error for scalar functions. The upper bounds on the error of approximating simultaneously all components of multivariable vector-valued functions are derived using a variable-basis approximation scheme in which all the scalar components share the same adjustable parameters inside the basis functions. The estimates are derived in terms of the Rademacher complexity of the families of functions. The advantages of simultaneous vector function approximation over componentwise approximation are discussed. Finally the results are applied to approximate dynamic programming.