|
|
|
|
|
|
|
An adaptive system faces its principal challenge when the set of possible structures  is very large and the performance functions µEinvolve many local maxima. It is important then for the adaptive system to provide itself with whatever insurance it can against a prolonged search. It is clear that the search of must go on so long as significant improvements are possible (unless the system is to settle for inferior performance throughout the remainder of its history). At the same time, unless it exploits possibilities for improved performance while the search goes on, the system pays the implicit cost of a performance less even than the best among known alternatives. Moreover, unexploited possibilities may contain the key to optimal performance, dooming the system to fruitless search until they are implemented. There is only one insurance against these contingencies. The adaptive system must, as an integral part of its search of , persistently test and incorporate structural properties associated with better performance. As with most insurance, this particular policy contains a limiting clause: useful properties must be identified to be exploited. The present chapter is concerned with this limitation. |
|
|
|
|
|
|
|
|
Almost by definition useful properties are points of comparison between structures yielding better-than-average performance. The question then is: How are the structures in a to be compared? If the structures are built up from components, comparison in terms of common components is natural and the question becomes: How is credit for the above-average performance of a structure to be apportioned to its components? A more general approach uses feature detectors (see section 3.4) to make comparisons. Since one can find an appropriate detector for any effectively describable feature of structures in (including the presence or absence of given components) this approach is well suited to present purposes. |
|
|
|
|
|
|
|
|
To begin with let us see how comparisons can be developed when a finite set of detectors  is given. In terms of the given detectors each structure  will have a representation (d1(A), d 2(A), . . ., d 1(A)); that is, each structure A will be described by its particular ordered set of l attributes or detector values  . Thus, for a chromosome A, Vi can desig- |
|
|
|
|
|