Re: Need ideas about building smarter agents

[Tony]

^Hello Paul,
^
^pauljohn> I wonder if anywone has the El Faroul simulation written for
^pauljohn> swarm?
^
^I also wonder the same :-) Please, let me know if you get some news.
^
^pauljohn> I'd like to build a model where agents can learn and adapt their
^pauljohn> strategies, and I'd really like some examples how to do it. Didn't
^pauljohn> someone in here talk about a stock market model with agents that 
learn
^pauljohn> (not like the one distributed with swarmapps?).
^
^Coincidently, after reading Arthur's paper I became interested in
^simulating the El Farol problem with some "realistic" learning
^agents. I found a reference that might be of your interest, "Modelling 
^Bounded Rationality using Evolutionary Techniques", B.Edmonds and
^Scott Moss. It can be found at
^
^http://www.cpm.mmu.ac.uk/~bruce/
^
^under "Publications".
^
^What would be the focus of your study on learning agents? The learning
^models themselves (so they ought to be as "realistic" as possible) or
^the overall collective behavior, with simple learning models? I would
^be glad to hear your ideas.
^
^Regards,
^
^Eiji
^

I think that one way of building learning agents is to make agents "nests" of 
lower level agents. So a low-level agent might be a strategy in the the El 
Farrol problem (eg, "Go if last week less than 40"), whilst a high level 
learning agent might be a collection of low-level agents linked by some 
updating process (eg a GA or GP). This allows the high-level agent to respond 
to changes in the environment by changing the frequency of low-level agents. 
High-level agents are separate to the extent that communication between them 
is strategic (ie not assumed to be costless and truthful). This is indeed the 
sort of approach taken by Edmonds and Moss. I also describe the approach in a 
09/97 Journal of Evolutionary Economics paper, an (old) draft of which can be 
found at:
http://price.econ.ucl.ac.uk/www/

Regards,

Tony

Tony Curzon Price, ELSE/UCL, London WC1E 6BT



                  ==================================
   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.
                  ==================================