Learning an optimal decision strategy in an influence diagram with latent variables

We consider influence diagrams with one decision variable in which some variables can be unobservable. Such a diagram is learnable if, under some regularity conditions (“strict positivity” ), it is always possible to find, with probability 1 and from an infinitely large sample, an optimal strategy for choosing the value of the decision variable. We give a simple graphical criterion for determining whether an influence diagram is learnable and show that an optimal decision strategy for a learnable diagram can always be found by a straightforward method. v Figure 1: An znfltience diagram wtth latent uartables