How Can a Bayesian Network Support Critical Infrastructure?

Welcome! What is a Bayesian Network?
May 11, 2019

One of the applications of Bayesian networks that very quickly illustrates their practical use is a decision support tool that has been encoded with expert knowledge for the use of non-experts. How is it “encoded with expert knowledge”? Well, talking with experts! A masterful facilitator (ahem) meets with experts in a particular domain to tease out the way they view the problem at hand, the different variables that play a role, how those variables are related and then assigning probabilities to different scenarios. Not going to lie, it can be a little tedious for the experts, as most of them don’t reflect on each observation, each step they perform, etc. It’s common to hear: “I don’t know….I just do it!” Eventually we get there.

One of the great things about Bayesian networks (IMHO) is how they can accommodate uncertainty. The world and the problems we face are riddled with uncertainty. In other decision tools, we are limited to binary options: If you observe X, they you should do Y. In a Bayesian network we use probabilities. In practice, it is rare that you definitely do Y if you observed X. Maybe if you observe X, it is 65% likely that the best option is Y. This is where Bayesian networks shine. Now stick with me, if I am losing you with the abstract “X” and “Y” stuff. The model illustrates this in practice.

This recently-developed demonstration tool considers the following scenario: There is a power outage in a cold and remote location. No experts are available to fix it. Is waiting the only option? Could a non-expert be deployed? Now, let me say up front, while this model was built by real experts in this field, it should not ever be used in a real-life situation. The model would only be part of the equation and contains decisions to be taken such as “replace transformer”. However, there are no instructions on how that would be done. This is strictly for demonstrative purposes. In other words, PLEASE DO NOT TRY THIS AT HOME!

Background

Above is a picture of a transformer. It “transforms” energy by stepping it down from a very high voltage (primary) that comes from the source of its generation to a lower voltage (secondary) that is distributed to homes and businesses connected via wires. While there are countless reasons the power can go out in a particular area; the first place to look is the transformer.

 

 

 

When power is lost, there is a switch that opens, essentially breaking the flow of electricity, and kills power going from the transformer.

 

 

Model

Here is our model! It was deliberately kept very simple. The arrows can be read as “causes me to update my beliefs about”.

The main thing I learned from the experts in this field is that they are incredibly practical people. Step 1, therefore, is….go look at the transformer!! If there is an obvious problem with it – the answer is also obvious – replace it. However, most often, even when there is something wrong with the transformer, it would not be clear just by looking, as demonstrated by the 96% likelihood that there is “No Evidence of Problem”. The model is a series of steps that should be taken based on either what is observed or the result of the previous step.

 

 

 

 

 

 

 

 

 

One of the features of the BayesiaLab software I like best is the use of Decision Nodes. These enable us to encode different expected values associated with different scenarios and then based on the inputs, it will tell us what the best action or policy is (light blue bar). The decision support tool is dynamic so when an observation has been made or an action taken, we tell the model that and it updates what we are likely to observe next or what we should do next.

 

 

 

This does not use any kind of machine learning or algorithm – purely expert insight. We basically develop a “virtual expert”. In this instance, we used number of man hours for our expected values for each scenario, which sometimes include a penalty, like “Doing nothing” in a situation where the only possible outcome is that people are out of power until “something” is done.

 

Imagine our non-expert has been deployed (obviously not you, since you are not attempting this!) and sees there is no evidence of anything wrong with the transformer (FYI, the experts in this field have all dealt with their fair share of electrocuted squirrels that resulted in said experts being pulled from their beds in the wee hours to remove said squirrel and replace the transformer. Unsung heroes!!). When we tell the network that, our “virtual expert” tells us that the best option then is to close the switch and see what happens. Either a fuse will blow or nothing will happen. It’s scary when the answer from your experts is “you wait for the explosion”. Let’s say that our non-expert does that and the fuse does indeed blow. We tell the virtual expert that and it will tell us the next best step to take.

 

 

 

 

 

We continue to take actions and observe outcomes until the problem is resolved. This type of tool can be available on hand-held devices so the non-expert just enters the actions taken and the observed results. While obviously not replacing an expert, this type of tools would, nonetheless, be hugely valuable in the event it is needed.

 

 

 

 

 

 

 

 

 

 

 

 

 

Putting it in Perspective

This kind of network build is labour-intensive, but certainly worth it when critical infrastructure is in question. Consider critical infrastructure at your company. What if your expert is on vacation? What if she is retiring and her valuable experience will be lost? What if your company is growing rapidly and your experts cannot keep up with the demand? Bayesian networks can be a user-friendly, intuitive tool to support critical operations.