[an error occurred while processing this directive] [an error occurred while processing this directive][an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive] (none) [an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive][an error occurred while processing this directive] [an error occurred while processing this directive][an error occurred while processing this directive] [an error occurred while processing this directive][an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive] (none) [an error occurred while processing this directive] [an error occurred while processing this directive] [an error occurred while processing this directive][an error occurred while processing this directive]
 
[an error occurred while processing this directive] [an error occurred while processing this directive]
Skåne Sjælland Linux User Group - http://www.sslug.dk Home   Subscribe   Mail Archive   Forum   Calendar   Search
MhonArc Date: [Date Prev] [Date Index] [Date Next]   Thread: [Date Prev] [Thread Index] [Date Next]   MhonArc
 

Re: [CPROG] Finde globalt max/min



Niels Langager Ellegaard wrote:
> Der er nogle folk der har brugt en masse fysik til at finde minima i
> deres neural netværk. De har tilføjet deres system en slags masse og
> en termostat (nose hover).

Nose-Hoover termostaten er den måde man typisk introducere en
temperatur i Molecular Dynamics (MD) simulationer. MD udføres
ved at numerisk integrere bevægelsesligningen og bevæger sig
derfor i NVE ensamplet (Fixed antal partikler, volumen, energi).
Ved at koble MD algoritmen til en stokastisk gnidningskraft i
bevægelsesligningen så kan man introducere en temperatur således
man sampler i NVT ensamplet istedet. Dette er tricket i Nose
Hover termostaten.

Monte Carlo simulationer har udfører typisk simulationer i NVT
ensamplet, fordi man har kontrol over temperaturen. Og simuleret
annealing er bare Monte Carlo hvor man gradvist sænker temperaturen
således at importance sampling gør at man sampler minimaer hyppigere.
(well i teorien, i praksis bliver bare bare fanget i et lokalt minima)

Mon ikke de har taget Back Propagation algoritmen, der er en
primitiv steepest descent metode og koblet denne til en termostat?

> Den er sikkert for indviklet til at være nyttig.
> http://eivind.imm.dtu.dk/staff/joe/nh.ps.gz
Lyder interessant, jeg vil kigge på den.

-- 
* Kurosawa: http://www.designlabs.dk/husetsbio *
* Email: Carsten dot Svaneborg at risoe dot dk * 
*  http://www.fys.risoe.dk/fys/External/casv/  *




 
Home   Subscribe   Mail Archive   Index   Calendar   Search

 
 
Questions about the web-pages to <www_admin>. Last modified 2005-08-10, 20:08 CEST [an error occurred while processing this directive]
This page is maintained by [an error occurred while processing this directive]MHonArc [an error occurred while processing this directive] # [an error occurred while processing this directive] *