To this point the major limitation of genetic plans has been their dependence upon the fixed representation of the structures . The object of the present chapter is to show how to relax this limitation by subjecting the representation itself to adaptation. This will be approached by reconsidering representation via detectors (chapter 4) in the light of the comment that detectors can be looked upon as algorithms for assigning attributes (section 3.4). Since algorithms can be presented as strings of instructions, the possibility opens of treating them by genetic plans, much as the strings of attributes are treated. (The mode of action of the genetic operators, of course, puts some unique requirements on the form and interaction of the instructions.) Actually, with a set of instructions of adequate power, we can go much further. We can define structures capable of achieving any effectively describable behavior vis-à-vis the environment. We can do this by setting up algorithms which act conditionally in terms of environmental and internal conditions. In particular, the predictive modeling technique of sections 3.4 and 3.5 can be implemented and subjected to adaptation. The Jacob-Monod "operon-operator" model (see the end of chapter 6) is suggestive in this respect, and we'll look at it more closely after the question of a "language of algorithms" (the instructions and their grammar) has been considered.
1. Fixed Representation
Before proceeding to a "language" suited to the modification of representations it is worth looking at just how flexible a fixed representation can be. That a fixed representation has limitations is clear from the fact that only a limited number of subsets of can be represented or defined in terms of schemata based on that representation. If is a set of structures uniquely represented by l detectors, each
. The object of the present chapter is to show how to relax this limitation by subjecting the representation itself to adaptation. This will be approached by reconsidering representation via detectors 
(chapter 4) in the light of the comment that detectors can be looked upon as algorithms for assigning attributes (section 3.4). Since algorithms can be presented as strings of instructions, the possibility opens of treating them by genetic plans, much as the strings of attributes are treated. (The mode of action of the genetic operators, of course, puts some unique requirements on the form and interaction of the instructions.) Actually, with a set of instructions of adequate power, we can go much further. We can define structures capable of achieving any effectively describable behavior vis-à-vis the environment. We can do this by setting up algorithms which act conditionally in terms of environmental and internal conditions. In particular, the predictive modeling technique of sections 3.4 and 3.5 can be implemented and subjected to adaptation. The Jacob-Monod "operon-operator" model (see the end of chapter 6) is suggestive in this respect, and we'll look at it more closely after the question of a "language of algorithms" (the instructions and their grammar) has been considered.