Hi
>> Agents are not fully autonomous, since they contextualized and even
>> possibly constrained by the environment, right?
>> If one attributes a degree of autonomy to a reactive agent, like in the
>> work on adjustable autonomy, giving the agent the chance of not
>> complying or deviating from the norm, wouldn't it suffice for the agent
>> to be autonomous?
>>     
>
> I would not, unless you replace "'chance of not complying' with 'decision of
> not complying', but this takes us back to cognitive agents.
>   
Does it not depend on what research aim is? In the review I posted 
yesterday, Paul Thagard argues that the book "underestimates the 
importance of explanation compared to prediction. In psychological and 
neuroscience, computation is used in prediction, but the primary role is 
in explanation by showing how postulated mechanisms can generate phenomena."

Now, some people would argue that - for many phenomena - prediction is 
as much 'explanation' as you're going to get. Newton said of his laws of 
gravity that "I have not been able to discover the causes of those 
properties, and I frame no hypothesis... it is enough that gravity 
really does exist, and acts according to the laws which we have explained."

In economics, Friedman argued for the same test: does your theory have 
predictive power? All hypotheses have assumptions - but according to 
Friedman, the conformity of those assumptions to 'reality' is NOT a test 
of the validity of the hypothesis: the test is predictive power only. So 
any theory that has a rational agent as its foundation can't be 
falsified on the grounds that people aren't like that.

The obvious response there is "but economics doesn't have any predictive 
power, you numpty. If their theories don't predict successfully, 
economists at the World Bank just tell the country implementing it that 
they didn't do enough, or did it too fast; certainly, the theory is 
sound. So it's never falsifiable, in reality.

But there is an important point here: a pollster is not interested in 
whether a model uses a crude stochastic process, or has finely grained 
AI agents.  They might say: 1. we know this type of area contains x type 
of person; 69% of the time they vote this way; so we should or shouldn't 
put our money there.  The research aim is quite specific - and doesn't 
need or want to ask WHY they vote this way, any more than NASA needs to 
know why gravity works in order to get things into space.

I'm also still a newbie to modelling, but one example: I did a very 
simple model of producers, workers and consumers. It needed no more than 
a set of for loops, and I ran it in Matlab. Now, it could be that I 
could 'get more' by giving each agent vastly more cognitive / 
computational complexity than "Pick a random place to purchase. Is this 
a price I'm willing to pay? Do I have the money?"  followed by changing 
price of product / labour accordingly.  But without actually explicitly 
stating what I'm trying to achieve, how would I know?  Without being 
explicit about this, we could play tennis all day with 'random / 
decision-making'. 

I'm not saying I do this yet, but I feel personally that I need to more 
clearly justify why I'd use AI agents, rather than using an aggregate 
approach, or just giving the agents mostly random choices - if these 
seem to approximate the behaviour I'm after.  Coming back to the first 
quote - "the primary role is in explanation by showing how postulated 
mechanisms can generate phenomena".  Ptolemy's geocentric theory of 
astronomy gave good enough predictions to be useful: it accounted for 
the movement of the heavenly bodies. So it met this goal of explanation 
- it showed a mechanism that generated what people saw and measured. But 
it did this through some fairly tortuous maths (and was wrong). What 
made Copernicus' system better? It was more parsimonious. It fitted the 
observations better, which is to say with less need for ad hoc add-ons. 
(Eventually, it got to be confirmed too, which is perhaps not something 
social simulators can do very easily.)

The point? It's quite possible to come up with more than one model to 
explain a phenomenon. Being still new to all this, I'm scrabbling about 
to find reasons why I should make things complicated when they could be 
simple.

Just to complicate matters further: Newtonian gravity has, in fact, been 
falsified by relativity. But it's still a perfectly good tool for most 
jobs that Earth-bound engineers need to do.  So again - and this is the 
key thing for me - *depending what your task is*, there's nothing wrong 
in principle with simplification.

I'm not sure where I'm going with this... thinking aloud. I'll stop now...

Dan