Re: Revised 10fcv results for set0

[Daniel Quinlan <qu...@pathname.com>]

Please use the -dev list for development discussion.  ;-)

I agree: 1,2,3.  Why?  I think TCR of 50 is probably still too low, so
I'm looking more at the FP rates.

- Daniel

Henry Stern <he...@stern.ca> writes:

> Rank by order of preference:
> 1:  0.0461%, 5.3556%, 11.6994
> 2:  0.0544%, 4.9949%, 12.0490
> 3:  0.0648%, 4.6278%, 12.2817
>
> I vote 1,2,3.
>
> No stat. sig. difference in TCR between 1,2.
> No stat. sig. difference in TCR between 2,3.
>
> There were no stat. sig. differences between the performance old 
> perceptron with the range_lo bug and the new one, so it must not have 
> been a big deal.
>
> Parameters:
> 1: -t 4.0 -p 2.0
> 2: -t 4.125 -p 2.0
> 3: -t 4.25 -p 2.0
>
> I'm going to generate the scores for all three and post them at
> http://stern.cs.dal.ca/sa-3.0/ [...] Will someone post whichever one
> that wins the vote to bugzilla and commit the scores/stats to SVN?
> [...]
> Henry
>
> P.S.  Don't peek at the statistics files in the gen-set0 directories 
> until you've cast your vote based on the 10fcv results.  There will be a 
> reasonable amount of variance in results on the test set between the 
> runs because each of the gen-set0 results are generated using a diffrent 
> random partitioning (and you'd only be looking at one sample!).

-- 
Daniel Quinlan
http://www.pathname.com/~quinlan/

Re: Revised 10fcv results for set0

[Theo Van Dinter <fe...@kluge.net>]

On Tue, Jul 20, 2004 at 11:18:50AM -0700, Dan Quinlan wrote:
> I agree: 1,2,3.  Why?  I think TCR of 50 is probably still too low, so
> I'm looking more at the FP rates.

Ditto on 1, 2, 3.  For set0, the FP rate is the main issue IMO.

-- 
Randomly Generated Tagline:
"You are confusing seriousness with excellence. You can pretend to be
 serious; you can't pretend to be intelligent. This is a common error
 among the technically unproficient." - "The Welcome Rain" on /.