%L eprints1763
%D 2007
%X Efforts to cope with the curse of dimensionality in Dynamic Programming (DP) follow two main directions: i) problem simplification by simpler models and ii) use of smart approximators for the cost-to-go and/or policy functions. Here we focus on ii). We consider: a) structural properties of the cost-to-go and/or policy functions (to restrict approximation to certain function classes and to chose the approximators); b) suitable norms of the approximation error (to estimate how it propagates through stages). These ingredients can be combined to develop efficient approximate
DP algorithms.
%A Giorgio Gnecco
%A Marcello Sanguineti
%A Riccardo Zoppoli
%C Prague, Czech Republic
%P 159
%T Exploiting Structural Results in Approximate Dynamic Programming
%O 22nd European Conference on operational research