From: Ankur Jain (ankur@cs.washington.edu)
Date: Wed Dec 01 2004 - 12:35:17 PST
This paper tries to empirically study if evolutionary theory can indeed
be used to explain the development of complex features such as eyes in
organisms; and hence put to rest some of the concerns that have been
raised about the theory in the past. For this, the authors trace the
evolution of a complex reward function in "digital organisms" which
started off without having any rewarding function. They make a couple of
interesting observations --
1) complex functions gradually evolved from several simpler functions
which served as their foundation.
2) even if simpler functions eventually got replaced by more complex
functions, it was essential that the simpler functions got rewarded when
they had appeared. If not, the complex functions did not get developed
at all.
From an AI perspective, their approach, in essence, boils down to a
local search of a very very large space with many good states. The
appearance of simple functions gives an indication of whether the search
is along the right path or not. Mutations provide for a mechanism to
make the search proceed. And just like local search, there might be some
moves that might in the short-run prove bad, but in the long run help
moving towards the final goal states.
While a very nice self-containing paper in itself, I was at loss
understanding how there experiments are very different from the genetic
algorithms that have for long been proposed in the AI literature. The
one significant point of departure could be that they show that
development of *complex features* is indeed possible within this model.
But isn't developing a complex feature akin to finding one of the goal
states which have that feature?
As an aside, and the authors mention this too, it would have been
interesting having a similar study on sexual reproduction. Also,
probably something closer to reality would be to have more than just a
discrete set of reward functions.
This archive was generated by hypermail 2.1.6 : Wed Dec 01 2004 - 12:35:18 PST