ON THE CLASS OF CLOSED DYNAMIC PROGRAMS

From MaRDI portal





The paper deals with general (GDP) and closed (CDP) dynamic programs when monotonicity and contraction conditions are satisfied. A method, which systematically generates cost functions for CDP and approximations GDP using CDP is developed. The GDP is defined by its state, cost, policy, loss and action spaces where the loss function is assumed to be monotonic and contractive. A dynamic program (DP) is called CDP if it is closed under certain monotonic contraction operators. A generalized policy-improvement algorithm for CDP is developed and its convergence is proved. This algorithm generates a sequence of \(\in\)-optimal costs (within CDP) which approximates GDP in only finitely many iterations. The concept of piece- wise DP is presented and a sufficiency condition for GDP to be piece-wise is derived. The paper also includes discussions of affine DP, partially observable Markov decision processes and a stochastic inventory model.











This page was built for publication: ON THE CLASS OF CLOSED DYNAMIC PROGRAMS

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3328311)