Symbolic AI systems are designed and programmed, rather than trained or evolved. In nature they are algorithmic, yet powerful. They work function under rules. Symbolic AI systems are typically confined to a narrow task such as chess playing, or theorem proving. Thus, they tend to be very fragile, and rarely are effective outside of their assigned domain; for example, a chess-playing program would not, if at all, perform as well diagnosing malaria as would a disease diagnosing expert system.
Artificial Neural Networks (ANN) are computational-cognitive models based on the structure of the nervous system.
The difference, or maybe an advantage of ANNs over expert systems is that they are trained rather than programmed.
They learn and evolve to their environment, beyond the care and attention of their creator. Although
expert systems are capable of pattern-matching and learning to some degree, the amount of learning that an
ANN can undergoe is greater as well as more flexible. For example, consider the knowledge that human being possess.
It would be quite impossible to straight-forwardly program a system that would store and manipulate
information to that capacity (you would have to specify everything a human being knew- manually). The problem would
be made much more feasible if we created a "learning machine". Learning is an important prerequisite for artificial minds.
ANNs are most widely used for pattern recognition or classification problems, however in theory, anything any computer
can do can be accomplished by an ANN.
Computational neuroethology (CN) is the study of a systems behavior within an environment.
It is concerned with the modeling of the behavior of these systems, as well as their neural substrates
(which is what neuroehtology is concerned with). CN systems percieve their environments directly. i.e.
they are not stored in some global database that was created through human input. They work in a
"closed-loop" environment, free from outside interactions. Their actions are solely based on what they
conclude from the state of their environment and as well as their prior actions. For example, if
we wish to simulate a robot in a closed-loop environment, then it must act not based on whatever semantics
or clues that could be provided from a human, but simply from the changes (or the state) of the environment
that it is in.
CN systems learn neurally and evolve genetically. They are adaptive, and act based on the circumstances
that they face in their environment. The drawback with CN systems are that they require enormous computational
resources. However, many CN advocators insist that much more can be gained from building complex models of simple
animals (systems) then from building simple models of complex animals (which is the traditional, and
more direct approach). Following this idea, the ultimate goal of AI is to create a human being. Yet to
accomplish this, we must first create a baby, not a full-grown adult.
Artificial Life (AL) is the study of artificial systems that behave like other natural living systems.
One example of AL is Reynold's Boids (link to Java adaption of Boids - Flozoids). This is a computer model
of flocking behavior in animals (such as birds or fish). One of its characteristics is that, the flock
(which is made up of many "boids") will
always reassemble if it passes through an obstacle (which causes it to scatter). How does it accomplish this?
Well, each boid follows a
few rules, such as: don't fall behind, keep up with nearby boids, try to stay a minumum distance
between your neighbors and obstacles, move towards what seems to be the center of mass of nearby boids.
While these rules may seem very simple, the result is a bunch of boids behaving like a real flock.
If you would like to learn more about boids, see the exclusive interview with Craig Reynolds
or Artificial Life.
Another example of AL are surprisingly, computer viruses. Computer viruses exhibit reproductive behavior, usually with
the intention of making trouble, emulating their biological counterparts.
Artificial life systems are mainly concerned with the formal basis of life. Its not important how an AL
system was created, but how it acts and behaves under its environment. They attempt to emulate lifelike
behavior. Many AL systems often evolve and coevolve, simulating evolutionary processes for
adaptation to an environment (luckily computer viruses can not evolve, hopefully this will
remain true for decades to come). However, there are limitations to these simulations; this is because of the simple
fact that everything in the physical universe can't be fully detailed. The more accurate the details (as well
as, the more the details) that are given in the simulation, the more likely it is to be successful.
