More DAI readings this week, this time on taking ideas from other disciplines and considering their relation to computer systems. I like this sort of study (in particular artificial life), but it's hard to do well. It requires integrity in one discipline while being able to apply ideas to another. In my opinion, the readings this week were fine for the integrity in their own field, but weren't quite there for applying the ideas to computer systems. This is not a condemnation - the papers were not trying to do that. It's up to us to apply ideas to our domain, though, and I'm not sure how good a job we did of that.
Two readings on insect behaviour from Jean-Louis Deneubourg's group. The reading on task differentiation in wasps seemed particularly interesting from a distributed systems point of view. The metaphor is to think of each wasp as an agent in a distributed system. Agents are almost identical, but have slight differences so that when pairs face each other in a simulated competition one usually dominates the other. If you have the task an agent performs be based on who dominates whom, you get a completely decentralized (pairwise interaction only) way of distributing tasks across the agent population. It's not at all clear how easy or useful this would be to do with artificial agents, but it's a nice idea.
The neatest idea from this group's work is stigmergy, where agent-agent interaction takes place through the environment ("mediated", we'd say). Juan pointed out that DAI blackboards are kind of stigmergic.
One particularly common form of environmental interaction takes the form of pheremones, trails agents leave behind to mark parts of the environment. Part of what makes pheremones work is that they evapourate, so the data is naturally self-expiring, and they're physically situated in an environment, so deixis is easy. These ideas have been applied some to distributed systems, for instance the well-known papers on pheremone trails for router congestion control.
Bob Axelrod's book "The Evolution of Cooperation" provided some classic game theory analysis in an artificial agent system. The main value of this work for our purposes is probably as a model for thinking of some of the complexity of agent-agent interactions. In particular, there's the idea that one agent's value (fitness) in a system cannot be isolated, but only understood in terms of its interaction with other agents. This isn't always true in systems - if agents don't interact at all, then their value is independent of the rest of the population. The reality in today's distributed systems is of course somewhere inbetween, but I believe as we see more agents deployed online we'll get more interagent interaction, and therefore more convolution of the fitness landscape.