Re: Agent complexity

[jalex]

Gary Polhill <address@hidden> writes:

> Does anyone know of two ABMs claiming to show the same phenomenon,
> but one having very much simpler agents than the other?

I'm not sure what the appropriate metric is for determining
simplicity, but I've got a couple models which intuitively have this
behavior.

They're spatial models of a simple form of the Nash bargaining game
(divide-the-dollar) where the agents live on a square lattice.  Agents
play divide-the-dollar with their neighbors (typically von Neumann,
Moore 8, or Moore 24, but any subset of Moore 24 can be chosen), and
change strategies based on some update rule.  

The simplest update rule is "imitate the best neighbor," with a coin
flip in case of ties.  Over 99.5% of the time, the population
converges to a state where everyone adopts fair division (which, I
would argue, says something about the evolution of norms of
distributive justice).

More advanced update rules give the same effect, up to a point.  If
agents adopt the strategy with the best average score (within their
neighborhood), fair division dominates.  The same thing occurs if
agents are willing to consider any strategy which did better than
them, with the probability of adopting that strategy proportional to
the "success" of that strategy.  I take it that "imitate the strategy
with the best average score" is slightly more complicated than
"imitate the best neighbor," and "imitate with probability
proportional to success" is considerably more complicated than
"imitate the best neighbor."

What's interesting, though, is that if agents use a best-response
update rule (the most complicated one I consider) , they do _less_
well than they would if they used a simpler one.

If anyone is interested, I've got some papers on this.  The papers
don't discuss the odd effects of the best-response rule, though.

Cheers,

Jason

-- 
Jason Alexander                   
Logic & Philosophy of Science      email:       address@hidden
School of Social Sciences          department:  http://hypatia.ss.uci.edu/lps
University of California, Irvine   phone:       (949) 824-1520 


                  ==================================
   Swarm-Modelling is for discussion of Simulation and Modelling techniques
   esp. using Swarm.  For list administration needs (esp. [un]subscribing),
   please send a message to <address@hidden> with "help" in the
   body of the message.
                  ==================================