coefficient ranges selection problem

Hello everybody, I am Ke Zhang, a graduate student from South China Normal University.It’s quite nice that this new discussion board allow me to successfully registered and to discuss SLEUTH with you.

I started with Sleuth since August in 2010, and meet with several problems from compiling.I get much help from the old discussion board and Dr. Clarke and an experienced user in China which also known from here.

With all your help, I have come to the calibration phase, but I am not quite understand how to slect the coefficient range.

First, let’s see:
from the instance of demo coarse calibration,sorted by the Lee Sallee metric, the range of Diff, Brd,Sprd,Slp,RG is 1,(1,25),1,(75,100),(1,50)
then the coefficient range for fine calibration is selected as :
breed {0 – 25,5}
spread {0 – 20,5}
slope {75 – 100, 5}
road gravity {0 – 50,10}

Then in the fine calibration phase, the range of Diff, Brd,Sprd,Slp,RG is 1,1,10,(75,95),(1,50)
then the coefficient range for final calibration is selected as :
dispersion {1 – 5, 1}
breed {1 – 5, 1}
spread {7 – 12, 1} why not spread {5 – 15, 2} ?
slope {75 – 93, 3} why not slope {75 – 95, 4} ?
road gravity {0 – 50, 10} why not road gravity {1 – 50, 9/10)?

Does the selection of coefficient range have some rules to follow?
or is it selected by feel? or practice?
(you know, a calibration process needs much time on my pc, nearly 48 h it has finished 58.6%)

The highest Lee Sallee metric value is 0.90904.I calculated the OSM by OSM=ComPare*PoP*edges*clusters*slope*Xmean*Ymean in Excel.
The highest OSM is 0.522936, with a Lee Sallee value of 0.90554.
This is my first fine calibration, and I use Lee Sallee metric to sorted the control_stats.log, select the range for fine calibration.

Would you recommend me to calibrate use Lee Sallee metric or OSM?

KE Zhang

One comment on “coefficient ranges selection problem

  1. The preferred way to do the calibration is to use the OSM. I will e-mail you the paper in which we developed and tested this metric. The fact that you are getting such a high Lee Sallee leads me to think that your calibration is biased toward low or zero growth. Obviously, no growth at all produces a Lee Sallee of 1.0, which is not very useful! These is a C program posted on the web, which calculates OSM directly from the log files. You could download and compile that, although using Excel works too.

    Ideally, by the final stage in calibration, it would be best to have unit increments. It is also best to limit the ranges carefully to minimize the number of runs necessary.

    You could also check out some new work I am doing with Matt Clarke-Lauer, on using Genetic Algorithms to do the calibration. When it is ready, this code will be placed onto SourceForge.

    Keith Clarke

Leave a Reply