|
|
|
|
|
|
|
correlation between attributes and performance, the set of schemata cuts through the space of structures in enough ways to provide a variety of "enriched" subsets. Intrinsic parallelism assures that these subsets will be rapidly explored and exploited. |
|
|
|
|
|
|
|
|
2. The "Broadcast Language" |
|
|
|
|
|
|
|
|
Though the foregoing is encouraging as to the range of partitions offered by a given set of schemata, something more is desirable when long-term adaptation is involved. First of all, when the payoff function is very complex, it is desirable to adapt the representation so that correlations between attributes and performance are generated. Both higher proportions of "enriched" schemata and higher "enrichment" factors result. It is still more important, when the environment provides signals in addition to payoff, that the adaptive plan be able to model the environment by means of appropriate structures (the  component of  in section 2.1). In this way large (non-payoff) information flows from the environment can be used to improve performance. As suggested in sections 3.4 and 3.5, by a process of generating predictions with the model, observing subsequent outcomes, and then compensating the model for false predictions, adaptation can take place even when payoff is a rare occurrence. |
|
|
|
|
|
|
|
|
To provide these possibilities, the set of representations and models available to the plan must be defined. Further flexibility results if provision is made, within the same framework, for defining operators useful in modifying representations and models. A natural way to do this is to provide a "language" tailored to the precise specification of the representations and operatorsa language which can be employed by the adaptive plan. Some earlier observations suggest additional, desirable properties of this language: |
|
|
|
 |
|
|
|
|
1. It should be convenient to present the representations, models, operators, etc. as strings so that schemata and generalized genetic operators can be defined for these extensions. |
|
|
|
|
|
|
|
|
2. The functional "units" (cf. detectors, etc.) should have the same interpretation (function) regardless of their positioning within a string, so that advantage can be taken of the associations provided by positional proximity (section 6.3). |
|
|
|
|
|
|
|
|
3. The number of alternatives at each position in a string should be small so that a richer set of schemata is provided for a given size of  (see the comments in the middle of chapter 4). |
|
|
|
|
|