Gigerenzer et al.

[M Lang / S Railsback]

Several months ago I asked whether anyone had read this book:
Gigerenzer, G., P. M. Todd and the ABC Research Group, 1999. Simple
Heuristics that Make Us Smart. There wasn't much response so in the
meantime I familiarized myself with the book and offer these comments.

The main topic is how individual agents (people, mostly, but animals are
also discussed) make decisions. Simple heuristics are offered as an
alternative to more conventional assumptions about how agents make
decisions, e.g., the "unbounded rationality" often assumed in economic
models. The heuristics have the obvious advantages of being much more
believable as a real decision process than assuming that people base
decisions on complex optimizations using perfect information. Simple
heuristics are claimed to be more realistic in the information and
computations needed to make a decision, and less subject to the risk of
"overfitting" than are statistical and optimization models. One goal if
the ABC Research Group appears to be developing a toolbox of ideas that
can be used to generate appropriate heuristics to individual problems;
they go out of their way to state that no single heuristic is likely to
be generally applicable to a lot of problems.

Most of the book consists of (1) identifying a particular type of
problem, (2) using Monte Carlo type analysis to compare the relative
success of simple heuristics vs. more complicated approaches in solving
the problem, and (3) exploring when and why the heuristics do well
compared to other approaches. Examples include:

1. The recognition heuristic: when attempting to identify the
alternative with highest value, choose the one you recognize. 

2. Take the best: when choosing among alternatives that each have values
for several different cues, (a) identify the cue that is most valid and
(2) pick the alternative with highest value for that one cue. 

3. To make a good choice from among a large number of alternatives,
examine a small number of the alternatives, then pick the next one that
has a higher value than any of those examined.

This book isn't likely to solve your modeling problems by identifying
just the right decision model for you, but it does provide a lot of good
ideas on modeling decisions. If nothing else, it identifies a number of
key issues that should be considered in finding a realistic model of how
people or animals make decisions- things like when an agent should stop
searching for more information. 

I was not always convinced by all the arguments- for example that simple
heuristics can outperform linear regression or Baysian statistical
models because they are less subject to overfitting. (Models were fit to
one data set, then tested against another set; statistical models that
fit the first data set too precisely are likely to do poorer in the
test.) Whether a regression model is subject to overfitting depends a
lot on the methods used to fit the model; there wasn't enough
information provided on the regression modeling methods to convince me
that the test was fair. Still, the authors do a commendable effort to
analyze the success of their methods quantitatively. (Agent based models
could be a great tool for comparing simple heuristics to alternative
decision modeling methods.)

I'm also not convinced that the animals I work with (fish) make all
their critical decisions in ways that can be modeled well with simple
heuristics. Many fish behaviors appear to be innate, evolved, responses
that seem too sophisticated to be modeled with simple heuristics. It
sounds funny, but simple heuristics seem more likely to apply to human
decisions about things like economics for which we are unlikely to have
evolved behaviors. 

Overall: If you are grappling with how to model decision-making by your
agents, then you are likely to find a perusal of this book interesting.
(If you have NOT thought much about how your agents make decisions, then
references like this are even more important!) I think it would be great
to get someone from the ABC Research Group to Swarmfest to update us on
their work.

Steve
-- 
address@hidden
Lang, Railsback & Assoc.
250 California Ave., Arcata CA 95521
707-822-0453; Fax 822-1868


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