Skip to main content

User login

What is not?

 Impossible is not a fact, its an opinion 

Certainty Factors -- Why do we need them?

May 23, 2010 by dangyogi

What is knowledge?  And how do we apply it?

Knowledge is knowing that a low fever is often associated with a common cold.

But what does "often associated" mean?  Do we have hard statistics on this?  Or is this more of a vague term?

Computers are simply machines.  They have no soul, no intellegence, no capacity for judgement.  They just blindly (and blindly is the important part) execute a series of instructions.  They have no concept that there is a purpose to executing these instructions.  No idea of what the goal is -- not even any understanding of the concept of a goal.  No way to constructively add their own capabilities (intelligence, experience, judgement) towards the task of reaching that goal; since they lack both those capabilities themselves, as well as any understanding of what the goal is.

So what we are trying to do is encode human knowledge in such a way that it can deposited into a computer and run through some blind mechanical process to arrive at a goal that we understand (and value), but the computer certainly does not.  We are not asking the computer to learn.  We are not asking the computer to develop its own experience or its own knowledge.

But we humans do have goals.  And we value those goals.  And if we can figure out how to encode knowledge and deposit it into a computer, it is relatively easy to replicate this knowledge a million fold.  And one of the interesting aspects of computers is that because they are strictly a mechanical process they will arrive at the same conclusion every time, given the same inputs.  They don't have bad days.

Computer hardware is built on electrical circuits.  These are simple digital circuits that are only capable of being either "on" or "off".  There are no inbetweens.  What's interesting is that several "on" or "off" circuits can be used together to form numbers (using a different numbering system than we humans are accustomed to -- called "binary").  So computers can process numbers.  That's what they are very very good at.  And numbers do have inbetweens -- the numbers between 1 and 10, for example, or 1 and 100, or 1 and a billion.

So the tool that we have in the computer to model vageness is simply numbers.  That's what "certainty factors" (CF) are.

We are interested in ariving at a diagnosis.  To do that we examine many pieces of evidence.  It is fairly easy to imagine attaching a number to each of these pieces of evidence.  The harder question is how do these numbers get combined to form a single number that represents the certainty that this diagnosis is the correct one?

That's what we have to figure out.  And it looks like there may be a half dozen different methods to do this; each with it's own reasons to use that method, as well as reasons not to use it.  The Pyke tool that we are using does not have any certaintly factor mechanism built into it.  The good news is that this allows us to develop each of these several methods and then to use each method where it makes sense to use it.  Or, stated differently, it doesn't look like there is one mechanism that will work for all diagnosis.

And figuring this out will require that we do many tests, to compare the results we get with this or that numeric method with the results that expert doctors produce.  These tests need to examine many different combinations of evidence to see whether the method works in all cases.

One last thing.  And I think that this is quite important.  Each piece of evidence (each question that we might ask) has three (not two!) different possibilities.

  1. The question is answered favorably towards reaching the diagnosis.
  2. The question is answered unfavorably.
  3. The question has not yet been answered.

So when you are considering how a question influences reaching a diagnosis, consider all three cases.  Our first CF mechanism, for example, makes no distinction between case 2 and case 3.  But is this always the right way to do it?

Premium Drupal Themes by Adaptivethemes