Compactification methods in a control problem of jump processes under partial observations (Q1277214): Difference between revisions

A control problem for a partially observed Markov jump process is considered. The process \(X_t\) takes values in a finite set \(I\) and has got its transition intensities \(\lambda(t,x,x',u)\) where \(0\leq t\leq T\), \(x,x'\in I\), \(u\in A\), \(A\) is a compact metric space. The observation process \(Y_t\) satisfies the equation \[ dY_t = h(t,X_t) dt + dW_t, \quad 0\leq t\leq T, \quad Y_0=0 \] where \(h\) is known, \(W\) is a Wiener process. The problem is to minimize the cost function \[ J(s,p,u) = E\left[ \int_s^T c(t,X_t,u_t) dt + g(X_T) \right] \] under control \((u_t)\) with values from \(A\) which depends on observations at the moment \(t\) for any time \(t \in [s,T]\). Here \(0\leq s\leq T\) and \(p\) is a law of \(X_s\). The following main results are proved: the existence of an optimal randomized or relaxed control; the equivalence of minima of cost functions under randomized and non-randomized controls; a dynamic programming principle. Also, a separation problem is considered and the existence of a Markov optimal filter is established.