Error in neldermead generated design

[miq]

Hi again - here is another problem I encountered, this time crashing NGENE:

Code: Select all

design
; alts = alt1, alt2, alt3
; rows = 16
; block = 4
? ; bseed=179424673
; rseed=179424673
?; bdraws = halton(1000)
; rdraws = sobol(100)
; eff = (rppanel,d)
; rep = 400
? ; orth = ood
; start = initial_ood.ngd
; alg = neldermead ( seed = 179424673 ) ? (converge=<float>, runs=<integer>, nointerim, seed=<integer>)
; con
; model:
U(alt1) = bsq[n,-1,0.5] /
U(alt2) = bA1[n,1,0.5]*A1[1,0]
        + bA2[n,1,0.5]*A2[1,0]
        + bA3[n,1,0.5]*A3[1,0]
        + bC[n,1,0.5]*C[-3:0:0.25] /
U(alt3) = bA1*A1
        + bA2*A2
        + bA3*A3
        + bC*C $


It starts running but at iteration 2, try 1000+ NGENE freezes and becomes unresponsive, and after a few minutes is closed.

[Michiel Bliemer]

You seem to have misspecified the continuous attribute levels. A continuous attribute level is specified something like C[-3:0}, where you specify the lower and upper bounds.
You have also specified a step, namely 0.25. This means that it attains levels [-3, -2.75, -2.5, -2.25, -2, .... , -0.25, 0], which is not continuous but discrete.
So please remove 0.25 as step, or replace neldermead with the normal swapping algorithm or mfederov algorithm, which are algorithms for discrete levels.

Michiel

[miq]

Hi Michiel,

Thanks for your reply. All clear.

I found this in the manual:

'Note that adding a second colon will generate discrete attribute levels, from the lower bound to the upper bound, with a step size specified after the second colon (e.g. [2:10:0.5]).' (p. 191)

and I thought the algorithm will choose the same number of levels as in the initial design, but choose whichever levels are most efficient from the candidate levels I specify via -3:0:0.25.

Thanks again!

Mik

PS. Is it useful to report cases that crash NGENE or generate unknown errors? Do you implement some measures to prevent these or should I just work around them?

[Michiel Bliemer]

Please report any instances that makes Ngene crash, as we investigate all of them.

In response to this crash we intend to build in a check that ensures that the input is consistent with the algorithm to avoid that Ngene crashes. Instead, it will generate an error message. This is hopefully in place for release 1.1.3.

So keep the crash reports coming please, they help us improve the software.