Re: How complex an agent?

[Pietro Terna]

        Dear Sven,

        it seems to me that the statements of the last message of Glen in this
list, related to the discussion opened by James Marshall, are very
important in your perspective.

        A guess: your interesting classification is mainly related to the 
software
design problem; the choice introduced by Glen (human cognitive constituent
functions vs. statistical - and I add: mathematical or formal - modeling of
human behavior) goes directly to the core of the agent construction problem.

        Pietro

At 14.15 08/11/98 -0600, you wrote:
>All,
>
>in designing the sim for my dissertation, I'm trying to come up
>with a gross classification of agent behavior, based on how complex it is.
>My advisor wants a 'full-circulation' model (i.e. an agent version of
>General Equilibrium), and I'm thinking that in such a model it may not be
>necessary that *all* agents be supersmart optimizing agents. Perhaps,
>depending on the question being asked, some of the agents can be of the
>'cardboard' variety -- just there doing their job to pass money and goods
>along as needed.
>
>(In fact, even in a full blown model we wouldn't want agents to be as smart
>as assumed in the usual rational-expectations economic model, since those
>models assume way too much knowledge on the part of agents.)
>
>It also seems like good software design ('stepwise refinement') to not
>create all agents as complex as possible from the start, since this makes
>it difficult to verify the model. (To verify the operation of one class of
>agent you want everything else to be as simple as possible, to avoid
>artifacts.)
>
>Anyhow, I offer up the list below of agent behaviors, in order of
>increasing capabilities, and invite your comments or criticism.
>
>Thanks!
>Sven Thommesen
>
>Level 0: agent performs the same act(s) every period [e.g. leaves the same
>amount of influence on the world grid], or responds the same way every time
>to messages from other agents.
>
>Level 1: as above, but response is contingent on variables in the agent's
>current state.
>
>Level 2: as above, but action/response is also contingent on data collected
>        a) from the agent's local neighborhood;
>        b) from globally available data.
>
>Level 3: as above, but give the agent MEMORY; it collects data (detail or
>summary) each period and current action/response depends on this memory data.
>
>Level 4: as above, but give the agent a GOAL, a function or criterion to be
>MAXIMIZED.
>
>Level 5: as above, but give the agent EXPECTATIONS over the future state of
>its world; many different expectations formation schemes are possible.
>Allow the agent to act so as to seek its goal (maximize its maximand) over
>expected futures.
>
>Level 6: as above, but add adaptation or a learning mechanism.
>
>
>
>
>                  ==================================
>   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.
>                  ==================================
>



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