Diagnosis and Troubleshooting
When your car doesn't start, what should you do?
Choice A) seems wiser. Why? Because checking your lights tells you much about your car's battery, a likely cause of your problem. But that's not the whole reason to check your lights first. Just as important, you check your lights first because such a check is easy and inexpensive.
MSBNx is a tool for doing this kind of cost-benefit reasoning for diagnosis and troubleshooting.
Causes and Uncertainty
You start using MSBNx by creating a model of your system. First, you create a diagram showing what causes what. For example, in our car model, the battery power causes the lights to on. Battery power also causes the engine to turn over which, in turn, causes the car to start.
Second, you tell about the uncertainty in the system, giving your probabilities for various situations. Here are some examples:
Third and finally, you can add cost information. In the car model, we set the cost of checking the lights at $1. We say that fuel pumps can't be tested, only replaced, which costs $100.
Your model of causes and effects is called Bayesian Network Network (or sometimes Bayes' Net or Belief Network).
Run It Backwards
Once you've created a model, you can use it for diagnosis and troubleshooting. MSBNx evaluates your causes-to-effects model by running it backwards from effects to causes. Given known effects, for example, evidence of the lights working, it computes the probability of causes such as a bad battery. It also computes the utility of further tests, for example, checking the gas gauge. Although, MSBNx recommends which test to do next, you can do tests in any order. As you enter new evidence, MSBNx efficiently recomputes the probability of possible causes and the recommended order of further tests.
An extensive manual documents MSBNx.
J. S. Breese and D.E. Heckerman. Topics in Decision-Theoretic Troubleshooting: Repair and Experiment. Microsoft Research Technical Report TR-96-06, 1996.
D. E. Heckerman, E. J. Horvitz, and B. N. Nathwani. Toward Normative Expert Systems: Part I The Pathfinder Project. Methods of Information in Medicine, 31:90-105, 1992.
D. E. Heckerman and J. S. Breese, et al., Troubleshooting Under Uncertainty. Microsoft Research Technical Report TR-94-07, 1994.
D. E. Heckerman and J. S. Breese. Causal Independence for Probability Assessment and Inference Using Bayesian Networks. Microsoft Research Technical Report TR-94-08, 1994.
E.J. Horvitz, J.S. Breese, and M. Henrion, Decision Theory in Expert Systems and Artificial Intelligence, Journal of Approximate Reasoning, Special Issue on Uncertainty in Artificial Intelligence, 2:247-302, 1988.