next up previous
Next: Experimental Setup Up: An Empirical Evaluation of Previous: Domains

Retrieval Strategy

Cases were initially retrieved on the basis of a static similarity metric which takes into account the goals that are covered by the case as well as all of their relevant initial state conditions [20,40]. Prior studies show it to be a reasonably effective metric. In learning mode, cases were also retrieved on the same basis. However, in this mode, the failure reasons attached to the case were used to censor its retrieval. Each time that a case was retrieved in learning mode, these failure conditions were also tested. If each failure reason was not satisfied in the new problem specification, the retrieval mechanism returned the case for replay. If, on the other hand, a failure reason was found to be true in the new problem context, then the case that repaired the failure was retrieved. Following retrieval, the problem was solved both by replay of the retrieved case as well as by planning from scratch.