large-n simulations

[Darold Higa]

Due to spam from my ISP, I have to repost this from a bounced post.
Thanks to Glen for helpoing me out!

>My current political-ecological model removes the "one agent per cell"
>restriction that many models use, and each agent I have contains a learning
>system (simplified AI).  With some careful programming (ie common lookup
>tables and the like) it is possible to have many thousands of agents that
>have some real computational ability.
>
>I can process simulations with agents well into the thousands.  Just don't
>expect real-time simulation runs.  My simulation has a population that
>varies between 200-2000 agents (as the population grows).
>
>I threw a machine together at home since the machine I began my Swarm
>project on was just too slow.  I used a single CPU 1GHz Athlon with 768MB
>RAM (not a very powerful machine) and I can still process about 200-300
>iterations in 12-24 hours, and a lot of the slowdown comes from the data
>collection process which writes about 8 rasters to disk per iteration (so I
>can make quicktime movies and animated gifs of the simulation runs).
>
>My agents feature a form of Holland's Learning Classifier System, so they
>analyze their surroundings, compare it to agent memory and then act.
>
>Hopefully I will have a website active soon with some results.  I agree
that
>distributed processing would be very useful, but if you do a lot of data
>collection, you can use slower machines to churn through the runs and then
>review the results at a later date.
>
>Also, if possible, program in the ability to turn on and off the display
>elements.  Most X-Windows managers are terrible resource hogs and a lot of
>memory and computational power goes there.
>
>It would be great if one day swarms and objects could be handled by
>different machines so that a single machine would actually handle any UI
>objects exclusively.  That way the simulation could proceed quickly with
>occasional interrupts from the "observerswarm" machine in order to grab
>display results for us slowpoke humans to observe the simulation results.
>This would be ideal for me as I have access to large numbers of machines
>that are not really fast.  I assume the biggest complication is all of the
>netcode that such a beast would require.
>
>Still, I don't think you need it if you are willing to not run in realtime.
>
>Darold Higa
>USC School of International Relations


                  ==================================
   Swarm-Support is for discussion of the technical details of the day
   to day usage of Swarm.  For list administration needs (esp.
   [un]subscribing), please send a message to <address@hidden>
   with "help" in the body of the message.