|
|
|
|
|
|
|
major respect: A broadcast unit can directly create other broadcast units, but a classifier, the broadcast unit's counterpart in a classifier system, cannot directly create other classifiers. This restriction permits a much simpler syntax based on only three atomic symbols, {1,0, # ("don't care")}. A classifier system creates new classifiers through the action of the genetic algorithm on the system's population of classifiers. |
|
|
|
|
|
|
|
|
Classifier systems aim at a question that seems to me central to a deeper understanding of learning: How does a system improve its performance in a perpetually novel environment where overt ratings of performance are only rarely available? A learning task of this kind is more easily described if we think of the system as playing a strategic game, like checkers or chess. After a long sequence of actions (moves), the system receives some notification of a "win" or a "loss" and, perhaps, some indication of the strength of the win or loss. But there is almost no information about what moves should have been changed to yield better performance. Most learning situations for animals, including humans, have this characteristican extended sequence of actions is followed by some general indication of the level of performance, with little information about specific changes that would improve performance. |
|
|
|
|
|
|
|
|
In defining classifier systems (see Figure 15), I adopted the common view that the state of the environment is conveyed to the system via a set of detectors (e.g., rods and cones in a retina). The outputs of the detectors are treated as standardized packets of informationmessages. Messages are used for internal processing as well, and some messages, by directing the system's effectors (e.g., its muscles), determine the system's actions upon its environment. Beside the interactions with the environment provided by detectors and effectors, there is a further interaction that is critical to the learning process. The environment must, upon occasion, provide the system with some measure of its performance. Here, as earlier, I will use the term payoff as the general term for this measure. |
|
|
|
|
|
|
|
|
The computational basis for classifier systems is provided by a set of condition/action rules, called classifiers. To simplify the computational basis, all interactions between rules are mediated by messages. Under this provision a typical rule, under interpretation, would have the form |
|
|
|
 |
|
|
|
|
IF there is (a message from the detectors indicating) an object left of center in
the field of vision
THEN (by issuing a message to the effectors) cause the eyes to look left. |
|
|
|
|
|
|
|
|
That is, the condition part of the rule "looks for" certain kinds of messages, and when the rule's conditions are satisfied, the action part specifies a message to be sent. Messages both pass information from the environment and provide communication |
|
|
|
|
|