Stochastic control methods in optimal design of life testing (Q1338752): Difference between revisions

We deal with stochastic control problems arising in the framework of accelerated life testing. These are experiments performed in order to learn about the qualitative behaviour of elements drawn from a population of identical individuals, without having to wait for a time comparable to the mean life of the element under test. We will analyze two specific models. In the first one we are interested in the failure rate of manufactured items: in this case the experiment amounts to submitting items under test to a stress level higher than the usual one. The second model makes reference to a population of living particles, and the stress is any action able to increase the fertility of the particle. In both models the experiment implies a cost, which typically increases with the test time, the level of stress, the number of elements under stress, and so on. On the other hand, increasing the values of such quantities is in general likely to improve the final information available from the test. Thus there is a trade-off, and a problem of optimal test design arises. Furthermore, we have to state how the stress acts on the evolution of the system. As it is usual in this kind of problems, the relationship between the stress level and the dynamics of the system involves some unknown parameters. Following the Bayesian point of view, we consider unknown parameters as random variables with a given initial distribution. The stochastic optimal control problem then becomes a problem with partial observation, the expected values of cost and return at each time have to be evaluated with respect to the current distribution of the parameters, conditioned on the collected data. This, in turn, requires the solution of a filtering problem. In order to transform the problem into one with complete observation (separated problem), we will include in the state the conditional distribution of the parameters with respect to the process observed up to time \(t\). We use probabilistic methods, that allow one to obtain very general results. They require use of weak controls where the probability space \(\Omega\) itself is a control parameter. In this context, by exploiting the special form of the generator of the controlled process, we prove that the optimal control problem is equivalent to a mixed control problem, where the number of particles under test is replaced by a stopping time. Finally, we provide, in a particular case, an algorithm allowing for the numerical approximation of the value function, and a discussion of the existence of sufficient statistics for the filter.