|
|
 |
|
|
|
|
(ii) An evaluator. The evaluator calculates an estimate of the "distance" of any given situation from the goal, using the detector outputs (an ordered set of real numbers) produced by that situation. The estimates are supposed to take the costs of the transformations, etc., into account; that is, the "distances" are usually weighted path lengths, where the paths involved are (conjectured) sequences of transformations leading from given situations to the goal. The intent is to use these estimates to determine which transformations should be carried out next. An evaluation is made of each of the situations which could be produced from the current one by the application of allowed (simple sequences of) transformations, and then that (sequence of) transformation(s) is executed which leads to the new situation estimated to be "nearest'' the goal. |
|
|
|
|
|
|
|
|
(iii) Error correction procedures. Before the search plan has been tried, the detectors and evaluator must be set up in more or less arbitrary fashion, using whatever information is at hand. The purpose of the error correction procedures is to improve the detectors and evaluators as the plan accumulates data. The shorter term problem is that of evaluator improvement. A typical procedure is to explore the search tree to some distance ahead of the current situation, either actually or by simulation, evaluating the situations encountered for their estimated distances from the goal. The evaluation of the situation estimated to be "nearest" the goal is then compared with the evaluation of the current situation and the evaluator is modified to make the estimates consistent. This "lookahead" procedure decreases the likelihood of contradictory distance estimates at different points on the same path. (A similar procedure can be carried out without lookahead using predictors to make predictions about future situations, subsequently modifying the predictors to bring predictions more in line with observed outcomes.) As a result, the consistency of the evaluator is improved with each successive evaluation. At the same time, in most searches, the difficulty of estimating the distance to the goal decreases as the goal is approached, becoming perfect when the lookahead actually encounters the goal. Thus increasing the consistency ultimately increases the relevance of the evaluator. |
|
|
|
|
|
|
|
|
There is, however, a caveat. If the set of detectors is inadequate, for whatever reason, the improvement of the evaluator will be blocked. This raises the broad issue of pattern recognition, for the set of detectors is, of course, meant to enable |
|
|
|
|
|