Re: [OctDev] regress and rcoplot (partially written, please help)

[Paul K. <pki...@us...>]

On Mar 9, 2006, at 1:49 AM, William Poetra Yoga Hadisoeseno wrote:

> On 3/8/06, Paul Kienzle <pki...@us...> wrote:
>>
>> I generally use only one digit for the tolerance.
>>
>
> OK, then use 1 digit :) The standard error values differ by as much as
> 10e-3 (as can be seen from the assert), and it doesn't improve if I
> embed the test in the main part of the script (that is, just using
> sqrt(v*c) as the standard error values).

Yet again I had to loosen the criteria on the tests in order to pass 
the tests on OS X PPC architecture.

The dataset is known to be hard, so I'm not surprised by the weak 
tolerances.  I'm curious if e.g., R can compute better results.

- Paul