Slashdot Mirror
Simulating Societies
blamanj writes "Most of us were exposed fairly early to Conway's game of Life.
A few simple rules produce a fascinating variety of behavior. Now, it
appears that similar simulations can predict the behavior of populations and human societies."
Psychohistory!
"Encyclopedia" is to "Wikipedia" what "Library" is to "Some people at a bus stop"
This article is really describing modeling using multiagent systems. Though very simple multiagent systems may resemble cellular automata (such as Conway's Life), they are not the same thing. Though they have been described in very convenient graphical representations using grids in the article, agents can model more complex behavior and need not be determinisitic (i.e. they may have a random element).
Another way to look at it is that cellular automata like Life use a single deterministic rule to govern the whole system. Agent-based systems, on the other hand, model goal-oriented behavior of the individual objects.
Again, Conway's game can be viewed as a very special case of an simple agent system, but the spirit of what is being done with agent systems is typically more involved. Comparing these systems to Conway's game of Life may create an incorrect impression for those not familiar with agent programming.
If you found this article interesting, their book is a great exposition of their early work with emergent behaviors. You can find it at Amazon here:
Growing Artificial Societies
There is a similar article on complexity and emergent behavior in the latest Harvard Business Review.
-XDG
Hmmm... So the simulation is accurate, but I would hypothesize that it does not show that a free society will trend towards "honesty."
Nonperiodic Central Trajectory
The problems lie elsewhere. Two that come to mind quickly are (1) lack of agreed upon factual data to use as the basis of the hypotheses. Do people with green skin have more or fewer babies out of wedlock than people with orange skin, and has this number increased or decreased over the last 10 years? Even in the US, with the Census data and tremendous amounts of market research, there are no agreed-upon answers to fundamental questions of data. Plenty of Newtons but no Kepler.
(2) None of these models are reversible. Put in a starting point of today's conditions, set the time increment to -1, and run the simulation backwards for 100 years. What comes out will be nothing like the world as it actually was in 1900. If we can't accuratly predict what happened in the past, how can we have any belief that the models tell us anything meaningful about the future?
sPh
You're correct about the models not necessarily being reversable - meaning that you can't predict history from the future. However, the correct method of verifying a simulation as correct is to verify the simulation results against known data. In the article, where it talks about the Anasazi, they describe writing the simulation and then letting it run through they years that they have data about the Anasazi (where the villiages are, the water availability, etc) and comparing it to reality. As described, they got quite close to reality. Villiages ending up in the same spot as reality over 50% of the time, etc. etc. etc.. Remember, it is very hard to determine the cause (or stimulus) from the effect without additional data. However, if the cause (stimulus) is known, the effect is usually fairly easy to guess.
If we were to try to build a model of today's history, you would want to build the model, seed it like the world was in the 1700's or earlier and let it run, and see how often it ended up correct. If it wasn't quite accurate, figure out where your model is wrong, fix, and repeat.
In the Asimov stories, what Hari Seldon was doing was to come up with a set of "formulas" (stored in the prime radiant) which accurately simulated history. The more accurate the formulas and the data you have, the more correct you are going to be. Hari and the members of the Foundation were constantly working on tweaks to better account for errors in the simulation. The hard part is dealing with the truly random influences. For instance, in the article when they talked about the Anasazi, they used real weather data instead of simulating it. I suspect if the weather data was simulated, the simulation would not have been as accurate on a year-to-year basis, although if the weather simulation was realistic enough I suspect that the outcome would have been similar.
Thinking back about Psychohistory as put forward by Asimov, I think that the only thing which really stretches for me is the accuracy (within a few months) of the events which he predicted-- taking into account the numerous variables which have such a rare occurance (such as an asteroid hitting a planet wiping everything out, or another major random event), that it would be throw the accuracy of small-scale events off. It seems logical that you can be accurate on a large scale on a simulation (over many thousands of years) or on a small scale (over a hundred years or so), but not both with the same simulation.
The problem with the "running forward from 1900" test is that the model includes, both explicitly and subconsciously, the model maker's view and understanding of the world that already exists. Including the events that occured between 1900 and 2000, say. So of course you would expect it to show reasonably accurate results for that time period - otherwise it would have been discarded during the development phase. However, that is no guarantee that the model is accurate outside the limits of that perception of the world.
I ran into exactly this problem myself. I developed several system dynamics models that seemed to give a good simulation of the population and wealth of the City of Chicago from 1950 to 1980. But when I ran them starting with the base data for similar cities, I got meaningless results. What seemed on first examination to be a general model of city population was actually just a condensed way of displaying the known state of one particular city.
So stronger tests than just "run forward to known state" are needed. Some argue that human events include irreversible processes, so perhaps the "run backwards" test is not valid. But more is needed than a demonstration between two known states.
sPh
/.
A fatal flaw of this simulation (as a model of real society, that is) is that it includes the "Cincinatus" characters - the incorruptible agents - but does not include the "Dillingers" - agents who are not deterred by punishment, of themselves or of others.
I have found over the years that people who are not influenced by "common sense" (or even an informed sense of self-preservation) are much more common than incorruptible people. Luckily (perhaps) these people more commonly are obsessed with greed than killing, or we'd have a lot more mayhem and a few less rich people.
Thus, the simultation should include agents that are not influenced by the arrest rate, and the model will probably become cyclic instead of trending to a fixed equilibrium.
Your statement that "the simulation is accurate" is unfounded, as any serious study of real behaviour in a police state will show. The Chinese shoot homosexuals and drug addicts; yet they still occur just as frequently as in other nations with less draconian laws. The US is "soft on crime" according to the Immoral Minority, yet our crime rates continue to drop.
But of course, anyone who thinks humans are simple agents with simple motivations is very unobservant.
--Charlie