Friam

A billion agents

Classic

List

Threaded

13 messages Options

Jochen Fromm-3

A billion agents

Besides the technical issues, what is the advantage of
parallel agent-based simulations ? Can you achieve more
with a billion agents than with a few thousand, or is it
just an attractive-sounding possibility ? An ant colony
with a billion ants will not be significantly different
or more intelligent than a colony with 10.000 ants. A swarm
with 10.000 birds will look similar to a swarm with 100
birds, only a bit more fine-grained.

Is a simulation with millions or billions of agents somehow
qualitative different from a simulation with only a few
thousand agents ? Certainly not if they are all alike, if
they all do the same or if they all "live" in the same
environment. I looks very difficult to construct a billion
different agents or to assign different tasks to billions
of agents.

In evolutionary systems, AI, and ALife, scale certainly
matters: a typical human brain has billions of neurons,
a chromosome contains roughly a GByte program with a
billion bytes, and evolution on Earth took from the
earliest forms to the computer nerd today a few billion
years. If we expect something interesting in an evolutionary
ALife system, do we have to let it run for some billion years
using a billion agents in order to get a "genetic code"
with a billion bytes ?

I bet the first true AI will have more than a billion bytes
of code, too (already a few films take easily a few GByte of
data). Somehow the lower bound for interesting behavior seems
to be a billion interacting units - why is this so ?

-J.

Russell Standish

A billion agents

Nobody knows until someone does the experiment. It is certainly
possible that something interesting will happen once enough agents are
simulated together. Right now it is a challenging task just to scale
the simulations up.

Cheers

On Mon, Oct 09, 2006 at 11:43:27AM +0200, Jochen Fromm wrote:

>
> Besides the technical issues, what is the advantage of
> parallel agent-based simulations ? Can you achieve more
> with a billion agents than with a few thousand, or is it
> just an attractive-sounding possibility ? An ant colony
> with a billion ants will not be significantly different
> or more intelligent than a colony with 10.000 ants. A swarm
> with 10.000 birds will look similar to a swarm with 100
> birds, only a bit more fine-grained.
>
> Is a simulation with millions or billions of agents somehow
> qualitative different from a simulation with only a few
> thousand agents ? Certainly not if they are all alike, if
> they all do the same or if they all "live" in the same
> environment. I looks very difficult to construct a billion
> different agents or to assign different tasks to billions
> of agents.
>
> In evolutionary systems, AI, and ALife, scale certainly
> matters: a typical human brain has billions of neurons,
> a chromosome contains roughly a GByte program with a
> billion bytes, and evolution on Earth took from the
> earliest forms to the computer nerd today a few billion
> years. If we expect something interesting in an evolutionary
> ALife system, do we have to let it run for some billion years
> using a billion agents in order to get a "genetic code"
> with a billion bytes ?
>
> I bet the first true AI will have more than a billion bytes
> of code, too (already a few films take easily a few GByte of
> data). Somehow the lower bound for interesting behavior seems
> to be a billion interacting units - why is this so ?
>
> -J.
>
>
> ============================================================
> FRIAM Applied Complexity Group listserv
> Meets Fridays 9a-11:30 at cafe at St. John's College
> lectures, archives, unsubscribe, maps at http://www.friam.org

--
*PS: A number of people ask me about the attachment to my email, which
is of type "application/pgp-signature". Don't worry, it is not a
virus. It is an electronic signature, that may be used to verify this
email came from me if you have PGP or GPG installed. Otherwise, you
may safely ignore this attachment.

----------------------------------------------------------------------------
A/Prof Russell Standish Phone 0425 253119 (mobile)
Mathematics
UNSW SYDNEY 2052 R.Standish at unsw.edu.au
Australia http://parallel.hpc.unsw.edu.au/rks
International prefix +612, Interstate prefix 02
----------------------------------------------------------------------------

Jochen Fromm-3

A billion agents

Yes, but it does not make sense to simulate
billions of agents just for their own sake.
I guess nothing interesting will happen until
each of the billion agents is unique and
contributes something different to the same
goal - which is only possible if the goal is
clear.

If the goal is the simulation of a city -
is a simulation of a whole city with 250,000
agents really different from a simulation
with 2,500 agents ? The reproduction of a
whole system in a ratio of 1:1 would be a
very poor model to understand it. And how
would you specify 250,000 agents with different
preferences ?

-J.

-----Original Message-----
From: Russell Standish
Sent: Monday, October 09, 2006 12:47 PM
To: The Friday Morning Applied Complexity Coffee Group
Subject: Re: [FRIAM] A billion agents

Nobody knows until someone does the experiment. It is certainly
possible that something interesting will happen once enough agents are
simulated together. Right now it is a challenging task just to scale
the simulations up.

Russell Standish

A billion agents

On Mon, Oct 09, 2006 at 01:23:29PM +0200, Jochen Fromm wrote:

>
> Yes, but it does not make sense to simulate
> billions of agents just for their own sake.
> I guess nothing interesting will happen until
> each of the billion agents is unique and
> contributes something different to the same
> goal - which is only possible if the goal is
> clear.
>
> If the goal is the simulation of a city -
> is a simulation of a whole city with 250,000
> agents really different from a simulation
> with 2,500 agents ? The reproduction of a
> whole system in a ratio of 1:1 would be a
> very poor model to understand it. And how
> would you specify 250,000 agents with different
> preferences ?

By setting their behaviour parameters from a probability distribution.

Also don't forget that geographical effects can be modelled in finer
and finer detail - a few thousand similar agents in a homogenous
environment probably is indistinguishable from bulk, but if you have
the street plan of New York, you may need hundreds of thousands of
agents before bulk effects come into play.

>
> -J.
>
> -----Original Message-----
> From: Russell Standish
> Sent: Monday, October 09, 2006 12:47 PM
> To: The Friday Morning Applied Complexity Coffee Group
> Subject: Re: [FRIAM] A billion agents
>
> Nobody knows until someone does the experiment. It is certainly
> possible that something interesting will happen once enough agents are
> simulated together. Right now it is a challenging task just to scale
> the simulations up.
>
>
>
>
>
> ============================================================
> FRIAM Applied Complexity Group listserv
> Meets Fridays 9a-11:30 at cafe at St. John's College
> lectures, archives, unsubscribe, maps at http://www.friam.org

Douglas Roberts-2

A billion agents

In reply to this post by Jochen Fromm-3

In the case of epidemiology simulations, much information is gained by
simulating all 300 million people in the US population to see patterns in
disease outbreak, and the relative effectivness of various intervention
strategies. Essential to this, of course is that the city/transportation
population networks and activity patterns be realistically represented.

--Doug

On 10/9/06, Jochen Fromm <fromm at vs.uni-kassel.de> wrote:

>
>
> Besides the technical issues, what is the advantage of
> parallel agent-based simulations ? Can you achieve more
> with a billion agents than with a few thousand, or is it
> just an attractive-sounding possibility ? An ant colony
> with a billion ants will not be significantly different
> or more intelligent than a colony with 10.000 ants. A swarm
> with 10.000 birds will look similar to a swarm with 100
> birds, only a bit more fine-grained.
>
> Is a simulation with millions or billions of agents somehow
> qualitative different from a simulation with only a few
> thousand agents ? Certainly not if they are all alike, if
> they all do the same or if they all "live" in the same
> environment. I looks very difficult to construct a billion
> different agents or to assign different tasks to billions
> of agents.
>
> In evolutionary systems, AI, and ALife, scale certainly
> matters: a typical human brain has billions of neurons,
> a chromosome contains roughly a GByte program with a
> billion bytes, and evolution on Earth took from the
> earliest forms to the computer nerd today a few billion
> years. If we expect something interesting in an evolutionary
> ALife system, do we have to let it run for some billion years
> using a billion agents in order to get a "genetic code"
> with a billion bytes ?
>
> I bet the first true AI will have more than a billion bytes
> of code, too (already a few films take easily a few GByte of
> data). Somehow the lower bound for interesting behavior seems
> to be a billion interacting units - why is this so ?
>
> -J.
>
>
> ============================================================
> FRIAM Applied Complexity Group listserv
> Meets Fridays 9a-11:30 at cafe at St. John's College
> lectures, archives, unsubscribe, maps at http://www.friam.org
>

--
Doug Roberts, RTI International
droberts at rti.org
doug at parrot-farm.net
505-455-7333 - Office
505-670-8195 - Cell
-------------- next part --------------
An HTML attachment was scrubbed...
URL: /pipermail/friam_redfish.com/attachments/20061009/f94f947e/attachment.html

Marcus G. Daniels-3

A billion agents

In reply to this post by Jochen Fromm-3

Jochen Fromm wrote:
> Is a simulation with millions or billions of agents somehow
> qualitative different from a simulation with only a few
> thousand agents ? Certainly not if they are all alike, if
> they all do the same or if they all "live" in the same
> environment.
It's not unusual to have semi-realistic landscape (e.g .from GIS data),
and then have agents move around on that landscape.
So even though the overall population is very large the effective
population acting on specific landscape over time is smaller. To get
good statistics on these smaller areas of interest, or various kinds of
rare events can mean that the whole simulation has to be very big.

Parks, Raymond

A billion agents

In reply to this post by Russell Standish

Russell Standish suggested that one could specify large quantities of
similiar but not exactly the same agents:

> By setting their behaviour parameters from a probability distribution.

But isn't this self-fulfilling? If you collect data about behaviours
to populate your probability distribution you will be programming your
agents to act the way you collected your data. If, by chance or design,
your data collection is biased, your agents will be biased.

--
Ray Parks rcparks at sandia.gov
IDART Project Lead Voice:505-844-4024
IORTA Department Mobile:505-238-9359
http://www.sandia.gov/scada Fax:505-844-9641
http://www.sandia.gov/idart Pager:800-690-5288

Marcus G. Daniels-3

A billion agents

Raymond Parks wrote:

> Russell Standish suggested that one could specify large quantities of
> similiar but not exactly the same agents:
>
>
>> By setting their behaviour parameters from a probability distribution.
>>
>
> But isn't this self-fulfilling? If you collect data about behaviours
> to populate your probability distribution you will be programming your
> agents to act the way you collected your data. If, by chance or design,
> your data collection is biased, your agents will be biased.
>

Being distributions, the parameters (the mixing ratios of different
kinds of agent behaviors) will have random peturbations around typical
values and in a large or long enough run you'll witness the consequences
of how this bias might play out at a global level.

The bigger the computers, the wider variances of agent mixes that can be
measured.

Douglas Roberts-2

A billion agents

Another way to answer this is, "No, it will not be self fulfilling if there
is an appropriate experimental design for using stochastically-generated
input parameters for agents in an ABM system." EpiSims uses stochastically
generated disease parameters to characterize both the disease agents and the
individual person responses to disease. When the EpiSims runs are made there
are additional stochastic processes that influence population mixing
patterns, with the results being statistically valid, and
non-self-fulfiling.

--Doug
--
Doug Roberts, RTI International
droberts at rti.org
doug at parrot-farm.net
505-455-7333 - Office
505-670-8195 - Cell

On 10/9/06, Marcus G. Daniels <mgd at santafe.edu> wrote:

>
> Raymond Parks wrote:
> > Russell Standish suggested that one could specify large quantities of
> > similiar but not exactly the same agents:
> >
> >
> >> By setting their behaviour parameters from a probability distribution.
> >>
> >
> > But isn't this self-fulfilling? If you collect data about behaviours
> > to populate your probability distribution you will be programming your
> > agents to act the way you collected your data. If, by chance or design,
> > your data collection is biased, your agents will be biased.
> >
> Being distributions, the parameters (the mixing ratios of different
> kinds of agent behaviors) will have random peturbations around typical
> values and in a large or long enough run you'll witness the consequences
> of how this bias might play out at a global level.
>
> The bigger the computers, the wider variances of agent mixes that can be
> measured.
>
> ============================================================
> FRIAM Applied Complexity Group listserv
> Meets Fridays 9a-11:30 at cafe at St. John's College
> lectures, archives, unsubscribe, maps at http://www.friam.org
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: /pipermail/friam_redfish.com/attachments/20061009/b3e6b889/attachment.html

Parks, Raymond

A billion agents

Douglas Roberts wrote:

> Another way to answer this is, "No, it will not be self fulfilling if
> there is an appropriate experimental design for using
> stochastically-generated input parameters for agents in an ABM system."
> EpiSims uses stochastically generated disease parameters to characterize
> both the disease agents and the individual person responses to disease.
> When the EpiSims runs are made there are additional stochastic processes
> that influence population mixing patterns, with the results being
> statistically valid, and non-self-fulfiling.
>
> --Doug
> --
> Doug Roberts, RTI International
> droberts at rti.org <mailto:droberts at rti.org>
> doug at parrot-farm.net <mailto:doug at parrot-farm.net>
> 505-455-7333 - Office
> 505-670-8195 - Cell
>
> On 10/9/06, *Marcus G. Daniels* <mgd at santafe.edu
> <mailto:mgd at santafe.edu>> wrote:
>
> Being distributions, the parameters (the mixing ratios of different
> kinds of agent behaviors) will have random peturbations around typical
> values and in a large or long enough run you'll witness the consequences
> of how this bias might play out at a global level.
>
> The bigger the computers, the wider variances of agent mixes that can be
> measured.

What I'm hearing is that you-all (or at least Doug) use statistically
valid distributions to program the behaviour of your agents. This seems
reasonable for things that can be measured with certainty - patient
response to disease, for example. However, it seems to me that if you
conduct a survey of a population sample concerning, say, political
opinion or purchasing habits, you run the risk of bias in the survey
being translated into bias in the ABM.

Let me suggest an example close to my own heart. I am interested in
how people react to retail market price changes in electricity in future
demand-response systems. There are some data from existing
demand-response systems but these function differently than newer
systems. What's more, the populations are geographically different from
the target populations and they are much smaller. We could survey
customers to see if they would allow their heating and air-conditioning
systems to respond to market prices. The trouble with past surveys of
this type is that utility customers tend to be more willing to sacrifice
comfort in the abstract than in reality. If one uses the PDF from the
survey, then the results are far different than the PDF from the
reality. And, the reality only works for certain locales and climes.

--
Ray Parks rcparks at sandia.gov
IDART Project Lead Voice:505-844-4024
IORTA Department Mobile:505-238-9359
http://www.sandia.gov/scada Fax:505-844-9641
http://www.sandia.gov/idart Pager:800-690-5288

Douglas Roberts-2

A billion agents

Interesting, Ray. A number of years ago a couple of LANL program managers
and I went back to Washington to try to interest DOE in an ABM that was
intended to do exactly what you describe blow. In addition, the simulation
would have allowed market gaming scenarios to be run for analyzing how the
power arbitragers (remember CalPine?) could game the market (remember all
those windfall profits the power companies were making in California a few
years back?).

We spent a few days, but basically could not find anybody at DOE willing to
give us the time of day, so the project never flew. We did mention a couple
of times during our visit that DOE does have the word "Energy" in it. Had
we gotten any support, the system would have been named ElectriSims, and
would have been another if the family of "SIM's" developed over the years by
our group (TRAMSIMS, EpiSims, MobiCOM).

--Doug (still completely unimpressed by the US Department of Energy --
rhymes with "Doh!")

> What I'm hearing is that you-all (or at least Doug) use statistically
> valid distributions to program the behaviour of your agents. This seems
> reasonable for things that can be measured with certainty - patient
> response to disease, for example. However, it seems to me that if you
> conduct a survey of a population sample concerning, say, political
> opinion or purchasing habits, you run the risk of bias in the survey
> being translated into bias in the ABM.
>
> Let me suggest an example close to my own heart. I am interested in
> how people react to retail market price changes in electricity in future
> demand-response systems. There are some data from existing
> demand-response systems but these function differently than newer
> systems. What's more, the populations are geographically different from
> the target populations and they are much smaller. We could survey
> customers to see if they would allow their heating and air-conditioning
> systems to respond to market prices. The trouble with past surveys of
> this type is that utility customers tend to be more willing to sacrifice
> comfort in the abstract than in reality. If one uses the PDF from the
> survey, then the results are far different than the PDF from the
> reality. And, the reality only works for certain locales and climes.
>
> --
> Ray Parks rcparks at sandia.gov
> IDART Project Lead Voice:505-844-4024
> IORTA Department Mobile:505-238-9359
> http://www.sandia.gov/scada Fax:505-844-9641
> http://www.sandia.gov/idart Pager:800-690-5288
>
>
> ============================================================
> FRIAM Applied Complexity Group listserv
> Meets Fridays 9a-11:30 at cafe at St. John's College
> lectures, archives, unsubscribe, maps at http://www.friam.org
>

Robert Holmes

A billion agents

On 10/9/06, Douglas Roberts <doug at parrot-farm.net> wrote:
>
> Interesting, Ray. A number of years ago a couple of LANL program managers
> and I went back to Washington to try to interest DOE in an ABM that was
> intended to do exactly what you describe blow. In addition, the simulation
> would have allowed market gaming scenarios to be run for analyzing how the
> power arbitragers (remember CalPine?) could game the market (remember all
> those windfall profits the power companies were making in California a few
> years back?).
>
<snip>

It's also interesting to note that many power arbitragers successfully gamed
the system without the benefit of massive simulations. It suggests to me
that we often get tied up in our own simulation/modeling expertise and
forget that there's a lot of practitioners out there with valuable domain
expertise. Give an arbitrager a solid financial incentive for finding
opportunities to game a system and s/he will find them a whole load faster
than a modeler :-)

R
-------------- next part --------------
An HTML attachment was scrubbed...
URL: /pipermail/friam_redfish.com/attachments/20061010/df9e77c5/attachment.html

David Breecker

A billion agents

One is tempted to say that the arbitragers are effective(ly) modelers ;-)
----- Original Message -----
From: Robert Holmes
To: The Friday Morning Applied Complexity Coffee Group
Sent: Tuesday, October 10, 2006 10:09 AM
Subject: Re: [FRIAM] A billion agents

On 10/9/06, Douglas Roberts <doug at parrot-farm.net> wrote:
Interesting, Ray. A number of years ago a couple of LANL program managers and I went back to Washington to try to interest DOE in an ABM that was intended to do exactly what you describe blow. In addition, the simulation would have allowed market gaming scenarios to be run for analyzing how the power arbitragers (remember CalPine?) could game the market (remember all those windfall profits the power companies were making in California a few years back?).

<snip>

It's also interesting to note that many power arbitragers successfully gamed the system without the benefit of massive simulations. It suggests to me that we often get tied up in our own simulation/modeling expertise and forget that there's a lot of practitioners out there with valuable domain expertise. Give an arbitrager a solid financial incentive for finding opportunities to game a system and s/he will find them a whole load faster than a modeler :-)

R

------------------------------------------------------------------------------

============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org
-------------- next part --------------
An HTML attachment was scrubbed...
URL: /pipermail/friam_redfish.com/attachments/20061010/4d2728a2/attachment.html