Russ (and everyone else),
Just because its what I know, I would do it in NetLogo. I'm not
suggesting that NetLogo will do what you want, just that it can
simulate doing what you want. Not knowing what you want to do, lets
keep it general:
You start by making an "agent" with a list of things it can do, lets
label them 1-1000, and a list of things it can sense, lets label them
A-ZZZ. But there is a catch, the agent has no commands connecting the
sensory list to the behaviors list, a different object must do that.
The agent must query all the rules until it finds one that accepts its
current input, and then the rule sends it a behavior code. (Note, that
any combination of inputs can be represented as a single signal or as
several separate ones, it doesn't matter for theses purposes)
You then make several "rules", each of which receives a signal from an
agent and outputs a behavior command. One rule might be "If input WFB,
then behavior 134." Note, it doesn't matter how complicated the rule
is, this general formula will still work. Any countable infinity of
options can be re-presented using the natural numbers, so it is a
useful simplification. Alternatively, imagine that each digit provides
independent information and make the strings as long as you wish.
Now, to implement one of my suggestions you could use:
1) The "system level" solution: On an iterative basis asses the
benefit gained by individuals who accessed a given rule (i.e. turtles
who accessed rule 4 gained 140 points on average, while turtles who
accessed rule 5 only gained 2 points on average). This master assessor
then removes or modifies rules that aren't up to snuff.
2) The "rule modified by agents" solution: Agents could have a third
set of attributes, in addition to behaviors and sensations they might
have "rule changers". Let's label them from ! to ^%*. For example,
command $% could tell the rule to select another behavior at random,
while command *# could tell the rule to simply add 1 to the current
behavior.
3) The "agents disobey" solution: Agents could in the presence of
certain sensations modify their reactions to the behavior a given rule
calls up in a permanent manner. This would require an attribute that
kept track of which rules had been previously followed and what the
agent had decided from that experience. For example, a given sensation
may indicate that doing certain behaviors is impossible or unwise (you
can't walk through a wall, you don't want to walk over a cliff); under
these circumstance, if a rule said "go forward" the agent could
permanently decide that if rule 89 ever says "go forward" I'm gonna
"turn right" instead.... where "go forward" = "54" and "turn right" =
"834". In this case the object labeled "rule" is still the same, but
only because the effect of the rule has been altered within the agent,
which for metaphorical purposes should be sufficient.
Because of the countable-infinity thing, I'm not sure what kinds of
thing a system like this couldn't simulate. Any combination of inputs
and outputs that a rule might give can be simulated in this way. If
you want to have 200 "sensory channels" and 200 "limbs" that can do
the various behaviors in the most subtle ways imaginable, it would
still work in essentially the same way, or could be simulated in
exactly the same way.
Other complications are easy to incorporate: For example, you could
have a rule that responded to a large set of inputs, and have those
inputs change... or you could have rules link themselves together to
change simultaneously... or you could have the agent send several
inputs to the same rule by making it less accurate in detection. You
could have rules that delay sending the behavior command... or you
could just have a delay built into certain behavior commands.
Eric
P.S. I'm sorry for the bandwidth all, but I am continuing to
communicate through the list because I am hoping someone far more
experienced than I will chime in if I am giving poor advice.
On Sun, Aug 23, 2009 10:32 PM, *Russ Abbott <[email protected]>*
wrote:
My original request was for an ABM system in which rules were
first class objects and could be constructed and modified
dynamically. Although your discussion casually suggests that rules
can be treated the same way as agents, you haven't mentioned a
system in which that was the case. Which system would you use to
implement your example? How, for example, can a rule alter itself
over time? I'm not talking about systems in which a rule modifies
a field in a fixed template. I'm talking about modifications that
are more flexible.
Certainly there are many examples in which rule modifications
occur within very limited domains. The various Prisoner Dilemma
systems in which the rules combine with each other come to mind.
But the domain of PD rules is very limited.
Suppose you really wanted to do something along the lines that
your example suggests. What sort of ABM system would you use? How
could a rule "randomly (or non-randomly) generate a new
contingency" in some way other than simply plugging new values
into a fixed template? As I've said, that's not what I want to do.
If you know of an ABM system that has a built-in Genetic
Programming capability for generating rules, that would be a good
start. Do you know of any such system?
-- Russ
On Mon, Aug 24, 2009 at 11:10 AM, ERIC P. CHARLES <[email protected]
<#>> wrote:
Well, there are some ways of playing fast and loose with the
metaphor. There are almost always easy, but computationally
non-elegant, ways to simulate things like this. Remember, we
have quotes because "rules" and "agents" are just two classes
of agents with different structures.
Some options:
1) The "rules" can alter themselves over time, as they can be
agents in a Darwinian algorithm or any other source of system
level change you want to impose.
2) The "rules" could accept instructions from the "agents"
telling them how to change.
3) The "agents" could adjust their responses to commands given
by the "rules" which effectively changes what the rule (now
not in quotes) does.
To get some examples, let's start with a "rule" that says
"when in a red patch, turn left". That is, in the starting
conditions the "agent" tells the rule it is in a red patch,
the "rule" replies back "turn left":
1) Over time that particular "rule" could be deemed not-useful
and therefore done away with in some master way. It could
either be replaced by a different "rule", or there could just
no longer be a "rule" about what to do in red patches.
2) An "agent" in a red patch could for some reason no longer
be able to turn left. When this happens, it could send a
command to the "rule" telling the "rule" it needs to change,
and the "rule" could randomly (or non-randomly) generate a new
contingency.
3) In the same situation, an "agent" could simply modify
itself to turns right instead; that is, when the command "turn
left" is received through that "rule" (or perhaps from any
"rule"), the "agent" now turns right. This is analogous to
what happens at some point for children when "don't touch
that" becomes "touch that". The parents persist in issuing the
same command, but the rule (now not in quotes) has clearly
changed.
Either way, if you are trying to answer a question, I think it
something like one of the above options is bound to work. If
there is some higher reason you are trying to do something in
a particular way, or you have reason to be worried about
processor time, then it might not be exactly what you are after.
Eric
On Sun, Aug 23, 2009 05:18 PM, *Russ Abbott
<[email protected] <#>>* wrote:
Thanks Eric. It doesn't sound like your suggestion will do
what I want. I want to be able to create new rules
dynamically as in rule evolution. As I understand your
scheme, the set of rule-agents is fixed in advance.
-- Russ
On Sun, Aug 23, 2009 at 8:30 AM, ERIC P. CHARLES
<[email protected] <#12349f6e507ad945_>> wrote:
Russ,
I'm probably just saying this out of ignorance, but...
If you want to "really" do that, I'm not sure how to
do so.... However, given that you are simulating
anyway... If you want to simulate doing that, it seems
straightforward. Pick any agent-based simulation
program, create two classes of agents, call one class
"rules" and the others "agents". Let individuals in
the "rules" class do all sorts of things to
individuals in the "agents" class (including
controlling which other "rules" they accept commands
from and how they respond to those commands).
Not the most elegant solution in the world, but it
would likely be able to answer whatever question you
want to answer (assuming it is a question answering
task you wish to engage in), with minimum time spent
banging your head against the wall programming it. My
biases (and lack of programming brilliance) typically
lead me to find the simplest way to simulate what I
want, even if that means the computers need to run a
little longer. I assume there is some reason this
would not be satisfactory?
Eric
On Sat, Aug 22, 2009 11:13 PM, *Russ Abbott
<[email protected] <#12349f6e507ad945_>>* wrote:
Hi,
I'm interesting in developing a model that uses
rule-driven agents. I would like the agent rules
to be condition-action rules, i.e., similar to the
sorts of rules one finds in forward chaining
blackboard systems. In addition, I would like both
the agents and the rules themselves to be first
class objects. In other words, the rules should be
able:
* to refer to agents,
* to create and destroy agents,
* to create new rules for newly created agents,
* to disable rules for existing agents, and
* to modify existing rules for existing agents.
Does anyone know of a system like that?
-- Russ
============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at
http://www.friam.org
Eric Charles
Professional Student and
Assistant Professor of Psychology
Penn State University
Altoona, PA 16601
============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at
http://www.friam.org
============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org
Eric Charles
Professional Student and
Assistant Professor of Psychology
Penn State University
Altoona, PA 16601
============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org
============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org
Eric Charles
Professional Student and
Assistant Professor of Psychology
Penn State University
Altoona, PA 16601
------------------------------------------------------------------------
============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org
