15may2002

Swarm intelligence


rough quote: "it's a mindset that consists of trying to solve problems from the
bottom up through self organisation and letting solution emerge"

four properties of a social insect colony:
- flexible
- robust
- decentralised
- self-organised

[what would Tron look like in a computer-as-swarm world?]

this is the problem: how do we shape emergence?

example of a solved problem: ants finding the shortest path between the nest and
the food point. but's it's not robust because they can get locked in to the
longer path if it's shown to the ants first. so with some species, the pheromone
trail degrades quicker, and that works better for switching paths.

and now he's applying this emergence thing to the travelling salesman problem.
you use ants dropping virtual pheromone, and there's maths on the screen and
everything. [damn. i want this maths and i can't note it down quite enough.
might have to ask for it.]

[i see. so this might not ALWAYS work, which is kind of what the biologic chap
was saying earlier.]

so this kind of algorithm doesn't nec get the OPTIMAL, but a very good solution
anyway. you end up with a bunch of nodes connected with varying amounts of
pheromone. follow it by always choosing the highest pheromone level at each
node. and this still works if connections between nodes are
appearing/disappearing a lot.

so this is a really good algorithm for routing, which is dynamic over very small
time scales.

[hey, this is really good, there's not much info to store: just an amount of
pheromone for each connection. which is just one column extra in dirk, for
example]

ooh: an agent updates routing tables on its way to its destination *viewing its
source as a destination*
and by having the pheromone evaporate, that makes it work better for dynamic
situations.

[!!] self organising communities: hyperlink connectivity

links to neighbors, jones is taking a photo. wow. [there's probability for
following hypertext connections, ways of reinforcing connections and laying
pheromones.] [AH! this is just the random surfer algorithm aka google's
pagerank]

[see also: that algorithm to cluster using pipes. so, by using this ant
algorithm you can figure out the pipe sizes for connectivity. and then the
pipe/flow thing (adding a node should flow out of the existing group less that
flows back in) does the clustering. wicked!]

[these notes don't really demonstrate how impressive this talk is. he's actually
showing demonstrations about how these algorithms work in the real world, AND
how much money could be saved using them for routing.]

he's now saying that these same algorithms can be used for sorting. ants do
scaling in many dimensions, and then project that data onto a 2d plane so you
can see where the clusters are.

[A few wrapup notes:
- there was a lot more in the talk than what i've just said
- unlike other talks, this wasn't nec a new metaphor system, it was a real way
of attacking problems. or maybe that's what a new metaphor system is?
- there were equations: I NEED THE SLIDES
- see also: the pipe flow algorithms in that issue of Computer [computer.org]
- dirk has 30000 connections, it's a good dataset to experiment and operate
on
- this is the kind of talk - this and the biologic one - that instead of
having to be just translated into my worldview, cf emergence, it actually
extends my worldview in new way.]

====
rumbling:

multitasking big style. im with people, talking with people, searching google,
listening to the talk. i'm getting better with practice doing this all together
too. continuously filtering, noting, looking up.

ah, actually. i'll wait a few weeks and read these notes to see how much better
i'm really getting.

snip of an AIM conversation:
etagolh: ooh. lemme dig out the wolf chapter from 'a thousand plateaux'. 
metadirk: is it online? 
etagloh: yeah, I think so. in one of the samizdat sites. 
metadirk: good grief. i'm talking with you, jones, listening to his and
searching google. i am the very nexus of the human hypertext. 
etagloh: damn. only in french. but: don't think ants, think wolves. 
metadirk: making emergence? hm. with hunting grounds? 
metadirk: 1730s. christ. 
etagloh: in terms of emergence, I'm thinking about predation, territorial
awareness, the sense of narrative, which is wolfish rather than antish.
certainly, you can code an ant, but you have to code as a wolf. 
metadirk: hm. i shall have to perculate. 
etagloh: my english d&g is in canada right now, but I will email you a ramble.