It is where they don their armor and charge, in the name of 10,000 euros.
"They" are pseudo-code strategies for a strange game that the organizers claim will shed some light on the computational properties of social behavior.
The rules of Project Cultaptation's Social Learning Strategies Tournament are complex, but the game is simple. Your strategy has to invade a population of 100 agents over thousands of simulation runs, based on its accumulated score. You get points by pulling levers of an n-armed bandit machine. But you can only pull a lever that you know about. And that knowledge can come from two sources: either you try your random luck, or you look at someone pulling levers next to you. Seems like a no-brainer, but the devil is in the details of this clever setup. Observing someone may be error-prone. The scoring may change over time, maybe faster than the expected lifetime of the agents.
I'm sorry to say that the deadline for entries to the tournament is over; it was today. Playing with the simulation has been truly fun. Of course, I really doubt that my entry will win -- I hope it gets past the round-robin selection phase! But I'd like to make a prediction about the winner.
I think the winning strategy will be very simple -- 10 lines of elementary pseudo-code maximum -- and stochastic -- with in-built randomness. I might have my reasons for that, but the chief ones come from history (the strategy that won the Prisoner's Dilemma tournament back in the 1980s was a very simple tit-for-tat) and perhaps misplaced reasoning-by-analogy (there's randomness in the environment, therefore a good strategy should include some of that too).
And I think the results won't tell us much about the fundamental nature of social learning, that is, gleaming truths about the way the world works from observing the behavior of others.
Why not? For instance, because in the tournament, my actions don't influence anything about the world beyond my own ability to score and therefore reproduce (ha ha). Or because there is no competition for resources -- all agents can choose the same action and they will all receive the same scores, same as it would be if only one agent chose it. Or even because there is no timeline, no sequences in this game, no planning a few steps ahead even.
Still, I can't wait to see the results, and they will be fascinating for all the simplifications of the game. And if I'm wrong, if the winning strategy is complex, or deterministic, or biologically plausible, it'll be pretty exciting to reason why that should be the case, and how the tournament actually captured some fundamental property of learning in groups.
Monday, June 30, 2008
Subscribe to:
Post Comments (Atom)
Posts
4 comments:
I had a real issue with the way they set up the rules. They made this completely unrealistic (for animals anyway) assumption that you couldn't recognize or remember your neighbors to evaluate how well they were doing. I think this is because some of the committee are in the business of publishing about how hard it is to evolve altruism, whereas in fact there are at least five different reasons that it evolves easily in species who disperse under their own power (walking / flying / swimming, in contrast to seeds getting blown around) and that live in mulitigenerational communities.
It looked like a fun hack, but I don't think with their weird memory restrictions the solution will wind up telling us that much about cultural evolution in insects and vertebrates. Sadly, I never got around to participating in the fun hack part, although I think one of my students did submit something (Marios Richards.)
Interesting perspective. I'm not sure that the observe/innovate problem is necessarily related to altruism. It's not like you get nothing if you innovate (not any less than if you observe anyway).
Also, the assumption that you can't recognize or remember neighbors could be warranted. For instance, apparently much of shoaling fish movement can be modeled with a set of very simple rules that don't assume any such ability. Can fish even recognize individuals, or only classify them as conspecifics vs. predators vs. prey? I'm not sure, but I've seen reported evidence that fish can observe others' behavior and learn socially, on some level.
Ah, but we agree the rules are too simple. Good luck to Marios.
Actually, apparently even guppies play tit-for-tat, so can recognize individuals. The fact that when you flock / school you generally follow simple rules doesn't mean that you might not accumulate info about your neighbours in other contexts.
Actually, Kevin Laland himself has a lot of social-learning in fish stuff, but is not the source of the tit-for-tat thing. He shows exactly that they follow their school in a maze, but they also learn, so if the old leaders are taken out the new fish that learned will still go the same way.
Yeah, ok, so apples and oranges. Tit-for-tat is not a valid strategy in the tournament because their game is emphatically not zero-sum (no sense of competition for resources). But schooling is a valid strategy, if it can be expressed as, say, "always observe". Or "observe more when younger" (an actual strategy that I submitted :-) ).
I guess that's exactly your beef with them, though. Fair enough.
Post a Comment