Mark, On 4/16/08, Mark Waser <[EMAIL PROTECTED]> wrote: > > > True, but this is inherent with ALL less than perfectly understood > systems and is not in any way peculiar to Dr. Eliza. Extrapolations are > inherently hazardous, sometimes without reasonable limit. > > Correct. Part of the point to AGI is to automatically create knowledge > bases that are as complete as possible. Dr. Eliza seems to be a reasonable > attempt to use a small amount of cherry-picked knowledge to solve a wide but > not complete range of unsolved problems of a given type -- and has all of > the standard inherent advantages and disadvantages of that approach. > Wouldn't you agree? >
Yes. > > There were a bunch of them and I don't claim to be a historian. As I > understood those methods they used two kinds of expertise - one of which was > similar to the symptoms and conditions that I use, and another that guided > the repair process. Dr. Eliza does without the guidance. This has the > advantage that it works with inept experts, and the disadvantage that it can > be less efficient than if it had good guidance. I had to find a grand > heuristic to replace expert-entered probabilities and the rest of that > guidance. After lots of experimenting, that grand heuristic turned out to be > incredibly simple, buried in the symptom weighting for various conditions, > being that you count the first potential symptom (or its verified absence) > as 80%, the next one as 80% * 20% = 16%, the third as 80% * 4% = 3%, etc. > This gives a lot of weighting to the leading symptoms, but nonetheless > seemed to work well. > > Wow! That's a *really* wicked tail-off. Seems really counter-intuitive. > Yes - it surprised me too, and it took a bunch of effort for me to get a good handle on why it worked, because I REALLY don't like my software to depend on things that I don't understand. It comes from Shannon's information theory. The amount of information in a datum is most dependent on the attendant noise. If you had a "perfect" symptom that exactly tracked a cause-and-effect chain link, then you would do best to ignore all other symptoms, regardless of whether they supported or contradicted the perfect symptom. In our less-than-perfect world, the list of potentially useful symptoms is usually short, and the "noise" comes from other cause-and-effect chain links that may exhibit substantially identical symptoms. If you have two symptoms, one with high noise and one with low noise, you do best by substantially ignoring the noisy symptom. The key to separating links using noisy symptoms is to use more than noisy symptom that hopefully has uncoupled noise. When your knowledge composer KNOWS about the 80% roll-off, then they CAREFULLY select which symptoms to use and which to ignore, for a secondary human effect of keeping the knowledge composer from "throwing in everything but the kitchen sink" along with the dirty wash water. Note further that unmentioned symptoms are NOT significantly considered in computing the result, only those that are affirmed or denied. This means that if ONLY the third symptom in the list that would only have a 3% effect if among others, has a 100% effect if it is alone. This results in noisy results - Dr. Eliza reports 100% interim probability, but fails to mention the >50% noise factor, and continues to press the user to answer questions about the two symptoms that precede the 3% symptom that is currently driving everything. Note also that the 3% symptom is probably also driving other potential conditions where it may be earlier in the list, and those conditions may are also be inserting their own questions. To separate the various 100%s in interim results, I added a heuristic to slightly reduce the 100% results proportionately to how far down the list that the first confirmed/denied symptom is. In typical use, there are often as many negative results (from denied symptoms) than positive results! What could a negative probability possibly mean? Not only do we have no believable evidence of the associated condition, but if natural forces were to try to force it, that those forces would probably fail approximately the indicated percentage of the time. > I'm not sure what you mean by "guided the repair process" > Where the expert's model of a decision tree, questioning, significance of symptoms, etc., is used instead of the engine's own generated one that may annoy the knowledge composer. It is interesting to watch others composing for Dr. Eliza, because they have their own ideas how to proceed in the presence of certain symptoms that may be of wide variance to Dr. Eliza's approach. So far, discussing this with them at length has yielded that that there really isn't any good reason for doing it their way, and by letting Dr. Eliza do its own thing, that inputting is a LOT easier. Note that there are NO expert-entered percentages in the Knowledge.mdb, which seems to result in BETTER operation because experts almost as often lead things astray with myths as guide things with overriding wisdom. Steve Richfield ------------------------------------------- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244&id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
