Thanks all for those who replied and gave the reference on various inference algorithms on QMR-like networks.
I was originally asking for performance of this kind of singly-connected networks in diagnosis, compared to the networks more complicated in structure, when they both model the same problem. The reason I am asking is because we can generate QMR-like networks with little effort. If in general they can give a satisfible diagnosis result, we do not have to allocate more effort to discover the dependency relations to build networks with a more complicated structure. The answer could be domain-dependent. However, I would like to see some theoretical analysis, and empirical comparison results in real applications. Thansk again, -Haiqin Wang Phantom Works, The Boeing Company