2003-04-23 Eric Swarm Intelligence: an example of bio-inspired computng Kevin Kelly: Dumb parts, properly connected into a swarm, yield smart results [ah, so that's the start of the swarm meme] great, but how to you connect the parts. things that do it: . brain . immune system . bacteria take inspiration form these social insects to do it wasps, termites, ants and bees -> create artificial insects: algorithms and systems to solve problems social insects: millions of years of evolution. they're very successful: flexible (internal perterbations, external challenges), robust (individuals may fail, tasks succeed), decetralised, self-organised the important thing in this presentation is the mindset decentralised/ bottom-up mindset how do we shape emergence [gardening!], how do we design individual behaviour to create this? demo of this and ants: "Double-bridge experiment" -- example of tipping between two paths to food by ants. the examples are embroidered [somebody used a word like that the other day in a talk. i really like it: it's just a matter of X, where X means embroidering, or complexifying, or something like that.], ants find the shortest paths. and again from last year: how to reintroduce robustness into the system using the technique of evaporation. get rid of the pheromones slowly. so this allows the system to adapt. although the simulation uses a faster evaporation rate than biology. now onto the travelling salesman problem. this is the same as last year... "a friend of mine says there are three things you don't do in public, and one of them in mathematics" [laughs] [same maths as last year, on the subject of how ants find the shortest path] modifying routing tables by pheromone. [yum] . agents are artificially delayed at congested nodes some big companies have used this method: . unilever plant scheduling . pina petroli truck routing . air liquide he says: interesting thing is this is a very short list section 2 bucket brigades in harvester ants georgia tech analysed this, and found it was the most optimal way of organising labour, so they also modelling the bucket brigade at Taco Bell okay, so it's to do with processes, how fast people go: it's all about queue theory. [there's some interesting queue theory docs about this, to do with removing bottlenecks and encouraging progress through processes in the NHS] he's discussing a game that's like this: the aggressor/defender game: two people, A is aggressor, B is defender. it's something to do about moving around keeping the "protector" between B and A. nothing happens when people play it. new rule: - you stand between A and B. everyone collapses to a single point between A and B. what happened? good news: you can simulate these systems to see what happens when people follow the rules, to see what the emergent behaviour is. [doesn't this depend on how robust the solution is to perturbations in the rules, regarding the change between simulation and humans playing the game as a perturbation?] but this worked for South West Airlines to optimise cargo routings: $10 million/ month savings A cautionary tale about blindly following simple rules: - oh, this is the thing about the army ants going round in circles until they die. Guyana observed a closed mill 1200 feed in circumference, circuit time of 2.5 hours for each ant. loads of ants died. "Rule #3 - no one needs to be in control" [embedded behaviour! you can't have anyone in control if everyone is localised] there's no-one in control when wasps built a nest. from simulations, eric has managed to make hive and nest-like structures with *only what can be perceived locally*. embedded bricks... [being able to *see* a distance is brilliant. it breaks localness, stops us being so embedded, although we end up being embedded in a different kind of system] something about system size vs system efficiency: explosive growth followed by saturation. eric says this is common in all kinds of emergent systems [so it's worth pursuing, see if there are ways to make use of this elsewhere]. bees solve the saturation problem by splitting the hive. ooh, nice bit by emergent behaviour by design. you can genetic algorithm-ly evolving the rules. "interactive evolution" [there are some simulations of social networks: people know who they hate and who they like. this is the problem with the other social networks we have: they do a lossy compression of the real world. it works for a simulation, but is it still useful in the other direction?] "Environment and wierd bugs in your simulation often end up having a big effect as well" [so the usual abstraction layer creation rules/methodology we use no longer works? we need to develop a new way, useful for something that has small changes making big effects] coming soon, the last slide: . controlling swarms . self aggregating devices . smart dust . social engineering of collecting phenomena . self healing comm networks Good question from Cory: . a "non-human readable world" where airflights get everyone where they need to go "on average" A: robust systems. . then cory rants again about the lack of transparency: how do you find out why a neural net has done something? There are loads of concerns about the army ants marching in circles. what if we end up following similar loops on much more complex and unidentifiable topographies? -> noise is good says eric [because it bumps you out of these sort of problems]