Gardenfors (2000)). But KBS deliver us no knowledge
of how knowledge takes place, whereas this is the goal
of BBKS.
4. Limits of BBKS
The proposed line of research is no panacea. We
can already see several limitations of this approach,
which should be considered if to follow this path.
When we model natural exhibitions of intelligence
(Figure 2), some people might say that we sin of
oversimplification, because we do not model all the
conditions which affect a natural cognitive system. But
simplification is a necessity, due to the immense
complexity of the phenomena which are modelled. It is
this complexity, and all the information (even when it
might be redundant) that our cells and brains can
contain, that force us to make simplifications in our
models. We believe that this information is so huge
that the complexity of natural organisms exhibiting
intelligent behaviour cannot be simulated in artificial
systems without simplification6. Our actual computers
are very far from being able to calculate in real time all
the necessary operations which a realistic (non
oversimplifying) model would require. Some
alternatives might lie in DNA computing (Benenson et.
al., 2001), but even if we had such computational
power as the one required to imitate convincingly
living organisms, how to program all the necessary
information? At this moment this seems impossible in
a short time scale.
But where to go? It seems we can make a
distinction depending on our purposes. If we are
interested in understanding intelligence (as we are with
the development of BBKS), then our limited models
creating artificial systems seem to suit our purposes. If
we want to produce intelligence “higher than human”,
we can learn a bit from the history of such attempts. In
the beginnings of Artificial Intelligence, some people
assumed that all the knowledge of a human adult might
be programmed. Other people aware of obvious
difficulties, looking at how natural systems acquire
their knowledge, thought of programming a “child”
computer that would be able to learn as a child does
(e.g. Turing, 1950) (of course, “one could not send the
machine to school without the other children making
excessive fun of it” (Turing, 1950)), (it is easier if you
do not program everything but let the parameters be
adjusted by the system, i.e. learned). Another
alternative has been to evolve the mechanisms in
charge of producing intelligent behaviour, also being
6This idea is clearly presented by Michael Arbib
(1989), speaking about brain models: “a model that simply
duplicates the brain is no more illuminating than the brain
itself” (p. 8).
inspired in nature (it is easier if you do not program
everything but let the model be adjusted by the system,
i.e. evolved), but it has required too much
computational power in order to aspire to reach
“higher order” intelligence by itself. These alternatives
and combinations of them have been used depending
on the ideas and purposes of researchers, modelling
from bacteria to human societies, all of them “sinning
of oversimplification”. So, if we want to produce
intelligence “higher than human”, it seems sensible
that we should not start building the computational
mechanisms from scratch. This is, we should not
attempt to throw away five billion years of evolution
and the computational power of our cells, and start
from where we already are, even from the hardware
perspective. This implies that we should build our
systems on us (The best model of a cat is another cat,
and if possible, the same cat). But for this of course we
need first to understand with our limited simplified
models how our mind works, in order to try to improve
it.
But we should notice that we already make
intelligence “higher than human”, just with our cultural
and technological evolution. There was already human
intelligence more than two thousand years ago, but we
could say that we are able to exhibit more intelligence
(we are able to solve more tasks) than humans of even
a hundred years ago (e.g. you can make calculations
much easierwith a computer than with pen and paper).
By altering the nature of our environments, changing
them to suit our purposes, we make our environments
and our tools to manipulate them more complex, and
our intelligence can be considered to be higher (Clark,
2003). Or from another perspective, we raise the level
of human intelligence, even with roughly the same
“hardware” (Our DNA has not changed much in the
last ten thousand years). Cultural evolution implies
that the intelligence will be improved each generation.
And the understanding of this process, will allow us to
guide it.
5. A Behaviour-based Knowledge
System
We are currently developing a BBKS in order to
study the development of knowledge in artificial
cognitive systems. Following the ideas presented in
Gershenson, Gonzalez and Negrete (2000), we are
constructing a virtual laboratory in order to contrast
our models as virtual animats develop and survive in
their environment. This virtual laboratory can be
downloaded (source code included) from
http://www.cogs.sussex.ac.uk/users/carlos/keb. A
screenshot of the virtual environment can be
appreciated in Figure 3.