|
|
|
|
|
|
|
situations, there may be no nontrivial homomorphisms.) The requirement on distinguishability permits the detectors to lump together situations having approximately equal expected utilities. This permits us to construct new, much smaller Markov chains based on events of interest. With this interpretation, the object of the adaptive plan tis to test different sets of detectors, H,and search plans, f. |
|
|
|
|
|
|
|
|
The error correction procedures for the detectors and the evaluator can be coordinated via a model of the search environment. In complex search environments, it is the detectors which make a model possible. From considerations of storage alone, a model would be impossible if, for each observed transformation, the initial situation and its successor had to be recorded in full. If the system records just the effect of transformations on the detector readings, it can be reduced to manageable proportions. Stated another way, the state space of the search environment is reduced to the manageable space of detector readings, and the effects of the transformations are observed on this reduced space. The construction of the model proceeds as data accumulate. When new detectors are required to increase distinguishability, an augmented model can be built around the old model as a nucleus. In particular, when a new detector is added to the set of detectors, this does not affect the data or the part of the model concerned with the old set of detectors. The task is to add information about the effect of the transformations upon the new detector, particularly in those situations that were causing difficulty. Once a model is available, it can be used with the evaluator to generate predictions and these can be checked against the outcome of that segment of the search. The resulting error indications, together with simulated lookahead, can then be used to improve the consistency of the evaluator. |
|
|
|
|
|
|
|
|
The sophistication of a model-evaluator search plan can only be justified when repeated searches must be made in the same overall search environment. As a prototype, we can consider an environment which (i) is complicated enough to make the exact recurrence of any particular situation extremely unlikely, but (ii) is regular enough to exhibit critical features (patterns) "pointing the way" to goals. The object of a plan, then, is to search out a succession of goals, improving its performance by incorporating the critical features in its detector-evaluator scheme. |
|
|
|
|
|
|
|
|
The overall objective of a formal study of this area is to find a plan which, when presented with any of a broad range of complex search environments, rapidly increases its search efficiency by extracting and exploiting critical features. Considered within the  framework, the domain of action  of a search plan  consists of the various combinations of detectors, models, and evaluators that the plan can generate. Usually these will all be specified as algorithms or |
|
|
|
|
|