Challenge:
If we compare either software systems, or robotic systems to biological systems, we find that our creations are incredibly fragile by comparison. Below we pose some challenges that are aimed at narrowing this distance. The challenges themselves are general in nature--they do not solve particular problems, but the acts of meeting these challenges will force the creation of new general purpose techniques and tools which will allow us to solve more particular problems.
Challenge 1.
Biological systems can adapt to new environments--not perfectly, they
die in some environments, but often they can adapt. Currently our programs are
very brittle, and certainly a program compiled for one architecture cannot run
on another architecture. Can we build a program which can install itself and
run itself on an unknown architecture? This sounds very difficult. How about a
program which can probe an unknown architecture from a known machine and
reconfigure a version of itself to run on the unknown machine? Still rather
difficult, so perhaps we have to work up to this by making some "blocks
worlds" artificial architectures where we can do this. This might lead to
some considerations of how future architectures might be designed so that
software is self-configurable, and then even perhaps self-optimizing.
Challenge 2.
In reality lots of problems we face in building real AI systems do not
get out of hand in terms of the size of problems for individual modules--in
particular with behavior-based systems most of the submodules need only deal
with bounded size problems. There are other ways theory could have gone. For
instance one might try to come up with a theory of computation based on how
much divergence there might be in programs given a one bit error in either the
program or data representation. If theory were based on this fundamental
concern we might start to understand how to make programs more robust.
Challenge 3.
Recent work with evolutionary system has produced some tantalizing
spectacular results, but it is hard to know how to take things from successes
and apply them to new problems. We do not have the equivalent of the Perceptron
book for evolutionary systems. (Note that these evolutionary systems are much
more that straight genetic algorithms as there is both a variable length
genotype and a morphogenesis phase that produces a distinctly different
phenotype.) We need such a new book of mathematics so that we understand the
strengths and weaknesses of this exciting new approach.
Challenge 4.
Can newer models provide us with new computational tools, and will they
lead to new insights to challenge the learning capabilities that we see in
biological learning?
Over time we become trapped in our shared visions of appropriate ways to
tackle problems, and even more trapped by our funding sources where we must
constantly justify ourselves by making incremental progress. Sometimes it is
worthwhile stepping back and taking an entirely new (or perhaps very old) look
at some problems and to think about solving them in new ways. This takes
courage as we may be leading ourselves into different sorts of solutions that
will for many years have poorer performance than existing solutions. With years
of perseverance we may be able to overcome initial problems with the new
approaches and eventually leapfrog to better performance. Or we may turn out to
be totally wrong. That is where the courage comes in.
Challenge 5.
Despite some early misgivings back when chess playing programs had
search trees only two deep (Newell et al. 1958), our modern chess programs
completely rely on deep search trees and play chess not at all like humans. Can
we build a program that plays chess in the way that a human plays? If we could,
then perhaps we could prove how good it was by getting it to play GO--tree search
just cannot cut it with GO.
Challenge 6.
All of the competitive speech understanding systems today use hidden
Markov models. While trainable, these systems have some unfortunate properties.
They have much higher error rates than we might desire, they require some
restriction in domain, and they are often inordinately sensitive to the choice
of microphone. It seems doubtful that people use HMM's internally (even if one
doesn't believe that generative grammars are the right approach either). Can we
build a speech understanding system that is based on very different principles?
Challenge 7.
We live in an environment where we make extensive use of non-speech
sound cues. There has been very little work on noise understanding. Can we
build interesting noise understanding systems?
Challenge 8.
Can we build a system by evolution that is better at a non-trivial task
than anything that has been built by hand?
4.jpg)
.jpg)
.jpg)
.jpg)
