From OpenCog
Jump to: navigation, search

On Fri, Dec 5, 2008 at 3:13 PM, Joshua Garner <> wrote:
> "5.10.1 Revision and Time
> Before proceeding with the mathematics, it's worth mentioning one conceptual
> issue that arises in the context of revision: the nonstationary nature
> of knowledge.
> If different estimates of the truth-value of an Atom were obtained at different
> times, they may have different degrees of validity on this account."
> I think this is the best short term approach and the one I will
> incorporate into Indefinite Truth Values.

Ben Goertzel replied:

If you want to keep time-stamped versions of TVs, then the way to do that is probably to use multiple TruthValue versions ...with a version corresponding to each time-stamp

On the other hand, the "degree of validity" notion is already incorporated into indefinite TVs of course ... it's there in the confidence, which is calculable based on the interval width U-L and the confidence level b ....

Creating the right equations for temporal decay of confidence is another issue ... it could be done by several ways such as

A) perhaps what you're suggesting: adding a **confidence decay factor** to each Atom, perhaps in the form of a DecayingIndefiniteTruthValue (extending IndefiniteTruthValue)

B) alternately... creating two kinds of decaying truth values, say...

-- RateSpecifyingDecayingIndefiniteTruthValue
-- RateDerivingDecayingIndefiniteTruthValue

The idea here is that an Atom with a RateSpecifiedDecayingIndefiniteTruthValue would actually contain a number specifying its truth value decay rate

On the other hand, an Atom with a RateDerivingDecayingIndefiniteTruthValue would not contain a decay rate number ... instead it would contain the Handle of another Atom, whose decay rate it borrows

The point of RateDerivingDecayingIndefiniteTruthValue would be to avoid having to manage too many decay rates ... when in fact most Atoms in the same general semantic category will share the same decay rate...

> I'm somewhat un-nerved by the talk of "confidence decay".
> When I read an authoritative  book on some new topic,
> I can be confident in what I read, and confident in what
> I know. I remain confident until I have some new information
> that challenges it. It would be a shame if confidence ebbed
> just because some time went by ...
> --linas

Yes, this is precisely why confidence decay is a subtle issue.

The rate of decay is context-dependent.

For mathematical knowledge, e.g., it is zero...

In an AGI system, the confidence decay rate must itself be learned
and adapted.  This can't be done for each Atom individual, so it has
to be done for higher-level contexts of knowledge.  This is a lot of
the reason I suggested the