Linear programming approximations for Markov control processes in metric spaces (Q1387656)

The authors develop linear programming techniques (LP) for discrete-time Markov control processes (MCP) in an uncountable metric space, particularly suitable for computational purposes. First, the authors introduce the associated LP equivalent to the MCP in the sense that \(\rho^*=\min\mathbb{P}\), where \(\rho^*\) and \(\min\mathbb{P}\) are the optimum values of MCP and LP respectively [cf. the authors' book, Discrete-time Markov control processes: Basic optimality criteria (1995; Zbl 0840.93001)]. Then, using a suitable aggregation-relaxation procedure to approximate \(\mathbb{P}\) by LPs with a finite number of decision variables, it is shown that under mild assumptions these procedures converge to \(\min \mathbb{P}\). The basic difference between the schemes of the authors and those of other researchers is that the authors use approximations in weak (or weak*) topologies rather than in normed topologies.