Greg Detre
Tuesday, April 08, 2003
Jordan Pollack, Brandeis
the
problem in AI is not simulating people, but how to write huge huge programs
Gould: argues that much of evolution is downward in terms of morphological complexity
life/alife � local reversal of entropy (i.e. creating organisation), dissipating energy/CPU cycles
second law of thermodynamics � cf knowledge/search tradeoff
No Free Lunch thereoms � machine learning is doomed � who???
Wolpert � violate this with coevolution
Brandeis DEMO lab
�
replicated Tesauro�s backgammon results with hill-climbing
TDGammon � one player played itself for 100,000 games
1 + 1 hill-climbing???
can print out robots - $50k plastic-printer � takes 24 hours
grammatical formalism called L-systems as the DNA
turtle language to build the bodies and to build the NNs
go here, turn right, add a joint etc.
add a neuron, connect it to this one � etc.
evolve the bodies and brains together
there were trying to build a dynamic oscillating kind of equilibrium
but it kept veering off into winner-takes-all
sometimes it can be more stable when you have more interacting populations, i.e. there�s a kind of critical mass for the ecology
mediocre stability
collusion
teacher-student game
what payoff matrix for the teacher and for the student will lead to the teacher asking hard questions that the student knows, or easy questions that the student doesn�t, which is the zone of best learning
why do they not want the teacher to only ask questions of difficulty 0.5 (as opposed to a normal curve with mean 0.5)???
�coevolution � it�s neither competition nor cooperation
with proper motivational structure � the right game � self-interested agents can provide each other with the information necessary to enable continuous learning�
Wolpert � collective intelligence
not bothered about the substrate � happy with both connectionism and selectionism
what�s the value of actually building the robots physically???
when they evolve recurrent nets, do they evolve the individual weights???
if not, is there any sense calling them NNs???
is there within-lifetime learning???
no � he thinks it�s faster to evolve fixed strategies
surely there has to be some external fitness involved even in coevolution though???
SS seems to think that the lesson is
no, and Pollack seemed to agree� L
you�re creating another level of
game
does coevolution act like a ratchet,
always moving towards improved fitness???