|
|
|
|
|
|
|
programs in some common formal language (see chapter 8). A search plan tÎ  thus amounts to a data-dependent algorithm for modifying the combination of detectors, model, and evaluator along the lines indicated above. The outward effect, at each point in time, of the combination produced by the search plan is a transformation h in the search environment. The range of the plan's action at each moment is therefore circumscribed by the set of possible transformations {h}. The set of admissible environments  consists of the set of search environments over which the search plan is expected to operate, each element E being presentable as a tree generated by the possible transformations. Let Ut,E(t) be the cost in E of the transformations applied by t through time t. If nt,E(t) is the number of goals achieved to time t, then Ct,E(t) = U,E(t)/nt,E(t) is the average cost to time t of each goal in E. A conservative measure of a plan's performance over all of e would then be |
|
|
|
|
|
|
|
|
 |
|
|
|
|
|
|
|
|
which yields the criterion  wherein plan t has a higher rank than plan t' if it is assigned a lower number by the above measure. Suggestions for a model-evaluator plan, based on the genetic algorithms of chapter 6 and capable of modifying its representations, are advanced in section 8.4. |
|
|
|
 |
|
|
|
|
Summarizing: |
|
|
|
|
|
|
|
|
 , [probabilistic] Markov chains induced by the sets of conditional probabilities {pi(S),the probability of applying transformation hi to situation  }; [general] admissible detector-evaluator-model combinations. |
|
|
|
|
|
|
|
|
W,[probabilistic] rules for modifying the conditional probabilities {pi(S)};[general] "lookahead" error correction, detector generation, and model revision procedures. |
|
|
|
|
|
|
|
|
, algorithms for applying operators from W to using information about the (sampled) average cost of goal attainment and (in the general case) errors in prediction ("lookahead") and observed inadequacies in detectors. |
|
|
|
|
|
|
|
|
e, the set of search environments characterized by search trees along with a transformation cost function µE(hi, S) giving the cost of applying hi in situation . |
|
|
|
|
|
|
|
|
, the ranking of plans in according to performance measures such as |
|
|
|
|
|
