Thunderhead Engineering Forum

Please login or register.

Login with username, password and session length
Advanced search  

News:

Forum moved to https://forum.thunderheadeng.com

Author Topic: random distributions  (Read 4157 times)

enronc

  • Member
  • **
  • Posts: 16
    • View Profile
random distributions
« on: February 01, 2010, 07:05:00 am »

Hi everybody,
I am working in evacuation simulations in tunnels with PathFinder.
I need the help of someone that knows about statistics.
My question is why I don't notice big differences about evacuation times (both in SFPE mode and STEERING mode) running different simulations of the same input files using normal distributions. I will explain it more clearly. The generation of a random normal distribution (about delay times, walking speed,etc.) generally causes variability in the results of the same input file due to the random "seed" that will be the start for the creation of the random distribution. So, in theory, when we create a normal distribution it is an artificial one that we create starting from pseudo-casual numbers (the used algorithm is generally the Box-Muller transformation). The software generally takes as starting seed the value of the clock of the pc, so it has to always change and all the distributions would change consequently and be different every time.

My questions are the following:

-  Which kind of seed does PathFinder use to create random distribution?
- Is the random normal distribution unique for each case (only one psudo-casual distribution is created)?
- Is possible to change in some way this distribution (maybe is possible to change manually the seed)?
- The solution I was thinking to solve this problem would be to create previously different normal distributions through the Box-Muller algorithm and then impute manully each value (but it will cost a lot of time!). Is there an easiest way to solve this problem in PathFinder?


Thank you in advance for your help and I am sorry if I have badly explained my problem!

Enrico
Logged

Charlie Thornton

  • Thunderhead
  • *****
  • Posts: 851
    • View Profile
Re: random distributions
« Reply #1 on: February 01, 2010, 09:10:42 am »

Enrico,

To re-distribute occupant parameters:

1. In the Tree View, right click Occupants
2. In the popup menu, click Randomize

This will change the seed (lots of them, actually) and cause all of the occupant parameters that use the random number generator to be recreated.  We intentionally designed the simulation to not use any random numbers, so the only values you are distributing are the occupant parameters (input).  We've also set up the input UI so that if an occupant has a particular generated value (e.g. initial delay), he will continue to have that value -- even if any other occupant is added or removed.  This makes it impossible to use a single random seed (one change would cascade though all remaining calculations).  This makes it so that your results will remain consistent between runs and we can also reproduce and fix any problems you may encounter that result from an otherwise hard-to-reproduce initial configuration.

Though, I admit, that randomize option is tough to find and it should be easier to run 50 randomized simulations and analyze the results.



Logged

enronc

  • Member
  • **
  • Posts: 16
    • View Profile
Re: random distributions
« Reply #2 on: February 01, 2010, 10:33:07 am »

Hi Charlie,
First of all, thank you. The randomize function is exactly what I was looking for. You know that the only problem will be that I have to launch 50 times the simulations to obtain the variability of the results for example in 50 runs. Have you thought about the possibility to implement in the next versions of PathFinder several runs in a row automatically?
Logged

Charlie Thornton

  • Thunderhead
  • *****
  • Posts: 851
    • View Profile
Re: random distributions
« Reply #3 on: February 01, 2010, 10:46:44 am »

Launching n runs for you would be easy.  It would also be super fast because we could launch as many simultaneous runs as you have CPUs.  Since the parallelism would be so coarse-grained you'd get linear speedup.

The part I haven't worked out is how we'll help you manage the data.  After we've dumped results in folders with names like mysim_1, mysim_2, ..., mysim_50, I'd like to be able to help you aggregate that data into useful results without having to paste it into excel one at a time.  Maybe I've been over-thinking it.  Should  we just make the runs easy and worry about it later?
Logged

enronc

  • Member
  • **
  • Posts: 16
    • View Profile
Re: random distributions
« Reply #4 on: February 01, 2010, 11:33:20 am »

Yes Charlie,
I have done what you have said to me and to run several simulations is easy and fast. The only thing is that later we have to copy and paste the obtained results on an excel spreadsheet.
Thank you,
Enrico
Logged

Charlie Thornton

  • Thunderhead
  • *****
  • Posts: 851
    • View Profile
Re: random distributions
« Reply #5 on: February 01, 2010, 02:11:35 pm »

Great, I'm glad you're making some progress.

By "we" I meant "we developers of Pathfinder."  Adding a feature that would launch n runs is something we could do, but I've been trying to figure out how to do it well.  When you have to launch n runs of Pathfinder, I imagine that's a boring, annoying, and time consuming process.

The question I was trying to get at was:

Should I just focus on making it easier to launch 50 pathfinder runs and leave the analysis to you?  or is it just as important that Pathfinder help to re-unite the data following the runs?
Logged

enronc

  • Member
  • **
  • Posts: 16
    • View Profile
Re: random distributions
« Reply #6 on: February 02, 2010, 05:06:56 am »

First, I think the best option will be to have the possibility to speed-up the process to obtain several runs in a row. The post-processing phase is something that every researcher makes in different way due to his specific needs. So it would be ok to have the evacuation results already in an easy-to-treat spreadsheet.
Logged