1 Introducing Bayesian Nets in AgenaRisk An example based on Software Defect Prediction

Typical Applications Predicting reliability of critical systems Software defect prediction Aircraft accident traffic risk Warranty return rates of electronic parts Operational risk in financial institutions Hazards in petrochemical industry

Typical Applications Predicting reliability of critical systems Software defect prediction Aircraft accident traffic risk Warranty return rates of electronic parts Operational risk in financial institutions Hazards in petrochemical industry

Typical Applications Predicting reliability of critical systems Software defect prediction Aircraft accident traffic risk Warranty return rates of electronic parts Operational risk in financial institutions Hazards in petrochemical industry

Typical Applications Predicting reliability of critical systems Software defect prediction Aircraft accident traffic risk Warranty return rates of electronic parts Operational risk in financial institutions Hazards in petrochemical industry

Typical Applications Predicting reliability of critical systems Software defect prediction Aircraft accident traffic risk Warranty return rates of electronic parts Operational risk in financial institutions Hazards in petrochemical industry

Typical Applications Predicting reliability of critical systems Software defect prediction Aircraft accident traffic risk Warranty return rates of electronic parts Operational risk in financial institutions Hazards in petrochemical industry

A Bayesian Net for predicting air traffic incidents

A Detailed Example What follows is a demo of a simplified version of a Bayesian net model to provide more accurate predictions of software defects Many organisations worldwide have now used models based around this one

Predicting software defects Operational defects The number of operational defects (i.e. those found by customers) is what we are really interested in predicting

Residual Defects Operational defects We know this is clearly dependent on the number of residual defects. Predicting software defects

Residual Defects Operational defectsOperational usage But it is also critically dependent on the amount of operational usage. If you do not use the system you will find no defects irrespective of the number there. Predicting software defects

Residual Defects Defects IntroducedOperational defectsOperational usage Predicting software defects The number of residual defects is determined by the number you introduce during development….

Residual Defects Defects found and fixed Defects IntroducedOperational defectsOperational usage Predicting software defects …minus the number you successfully find and fix

Residual Defects Defects found and fixed Defects IntroducedOperational defectsOperational usage Obviously defects found and fixed is dependent on the number introduced Predicting software defects

Residual Defects Problem complexity Defects found and fixed Defects IntroducedOperational defectsOperational usage The number introduced is influenced by problem complexity… Predicting software defects

Residual Defects Problem complexity Defects found and fixed Defects Introduced Design process quality Operational defectsOperational usage ….and design process quality Predicting software defects

Residual DefectsTesting Effort Problem complexity Defects found and fixed Defects Introduced Design process quality Operational defectsOperational usage Finally, how many defects you find is influenced not just by the number there to find but also by the amount of testing effort Predicting software defects

A Model in action Here is that very simple model with the probability distributions shown

A Model in action We are looking at an individual software component in a system

A Model in action The prior probability distributions represent our uncertainty before we enter any specific information about this component.

A Model in action So the component is just as likely to have very high complexity as very low

A Model in action and the number of defects found and fixed in testing is in a wide range where the median value is about 20.

A Model in action As we enter observations about the component the probability distributions update

Here we have entered the observation that this component had 0 defects found and fixed in testing

Note how the other distributions changed.

The model is doing forward inference to predict defects in operation…..

..and backwards inference to make deductions about design process quality.

but actually the most likely explanation is very low testing quality.

…and lower than average complexity.

But if we find out that the complexity is actually high…..

then the expected number of operational defects increases

and we become even more convinced of the inadequate testing

So far we have made no observation about operational usage.

If, in fact, the operational usage is high…

Then we have an example of a component with no defects in test..

…but probably many defects in operation.

But suppose we find out that the test quality was very high.

Then we completely revise out beliefs

We are now pretty convinced that the module will be fault free in operation

…And the ‘explanation’ is that the design process is likely to be very high quality

A Model in action we reset the model and this time use the model to argue backwards

A Model in action Suppose we know that this is a critical component that has a requirement for 0 defects in operation…

The model looks for explanations for such a state of affairs.

The most obvious way to achieve such a result is to not use the component much.

But if we know it will be subject to high usage…

Then the model adjusts the beliefs about the other uncertain variables.

A combination of lower than average complexity…..

…Higher than average design quality…..

and much higher than average testing quality …..

But suppose we cannot assume our testing is anything other than average…

Then better design quality …..

..and lower complexity are needed …..

But if complexity is very high …..

…Then we are left with a very skewed distribution for design process quality.

What the model is saying is that, if these are the true requirements for the component then you are very unlikely to achieve them unless you have a very good design process

Making better decisions That was a simplified version of model produced for Philips Helped Philips make critical decisions about when to release software for electronic components 95% accuracy in defect prediction – much better than can be achieved by traditional statistical methods

Model Implementation In AgenaRisk