Re: [Crm114-discuss] New Results (Re: CRM114 Testing: Urgent: Please reply)

[Fidelis A. <fi...@po...>]

On Sun, 18 Apr 2004 22:10:11 -0400, Bill Yerazunis wrote
> Fidelis:
>
> *fascinating*....
>
> I was talking today with Penney (my better half, etc) about how to trim
> the feature space down enough so that it could be the inputs to a neural
> net.  Right now there's a million features, and that's far far too many.
>
> I thought that getting rid of the intermediate features would cause
> too much loss of accuracy.  There just isn't space in a feature slot
> of reasonable size to do what's needed.
>
> But if you really _can_ ignore a lot of the polynomials, then that
> drops the dimensionality down enough to be reasonable as the first stage
> of a Hopcroft-style (feedback-capable) neural net.

Oh, neural nets... 18 years ago I was getting my master degree in AI and
used to play with prolog, lisp, natural language understanding, etc. At 
that time, some co-students were playing with neural nets and I was very
impressed with the possibilities, but life took me to other ways and I
had to concentrate on my job as a telecomm engineer and later as an 
internet servers support engineer. It's interesting to hear about neural 
nets again.

>
> And _that_ will kick butt.  :)

I hope so :)

>
> -----
>
> Now, as to the difference between 56 and 70, I have no clue.  But
> your set shows that you have a large standard deviation per run; I
> suppose I hit a good set.
>
> Did you try to replicate the run with the ten shuffles I sent out (or
> did I send them to you at all?  Maybe not...)

No, you didn't. But I'd like to run the test with it.

> But your results vary all the way from 62 to 82 out of a 10-shuffle set.

I believe that it is an indication that the learning step is greater
than necessary, which may cause unlearning of things already learned and 
  greater oscillation/delay till convergence, if it converges at all. 
Some sequences may be more sensible to that effect.

>
> Did you use the defaults for the build (same :lcr:, same default number
> of slots in the .css files?)

Yes, I used all the defaults. The only change, applied to the WPCW
version, was smaller css files. I used the prime 282001 for the number
of buckets - near 1/4 of the default size - because WPCW produces 4 
times less features than SBPH.

>
> Fascinating, either way though.

I'm also excited about the possible results... :)

--
Fidelis Assis