Re: [agi] fuzzy logic necessary?

[Yan King Yin]

 

I tend to agree with Richard's view and I may build an AGI with symbolic, non-numerical inference.

 

1.  As Russell pointed out, if the priors are not known or are in extremely low precision, Bayes rule is not very applicable.  Number crunching with priors of 1-2 bits precision is "garbage in, garbage out".

 

2.  It seems that in the majority of situations, priors are of 1-2 bits precision.

 

3.  To put it another way, it seems that in most real situations, the quality of an inference is often greatly improved by taking into account more facts / contexts, much more so than by increasing the precision of probabilities of a smaller number of facts.

 

4.  Even worse, this seems to be a fundamental feature of reality.  Can an AGI increase the precision of its internal probabilities by continually updating them?

 

5.  The answer seems to be negative.  Real events are dependent on a lot of other events in a complex way.  They usually do not repeat again and again under the same conditions.

 

6.  Priors can be known with great precision in 2 cases: 1) the outcomes are enumerable and equi-probable, as in rolling a dice; 2) the event have repeated many times under identical conditions.  It seems that the majority of events are not like these.

 

YKY

---

To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]