In our previous research, we investigated the properties of case-based plan recognition with incomplete plan libraries. Incremental construction of plan libraries along with retrieval based on similarities among planning situations (rather than on similarities among planning actions) enables recognition in light of novel planning actions. In this paper we investigate the recognition behavior in situations where the recognizer fails to find past situations that match the currently observed situation at any level of abstraction. Such recognition behavior is especially common in early recognition stages when the rate of new bin observations is large. To cope with newly observed situations, we employ a retrieval scheme that utilizes a similarity measure among the states in the abstract state-space, based on the k-nearest neighbor similarity metric. Such a retrieval scheme may enable recognition in light of newly observed abstract situations. Properties of the retrieval in abstract state-spaces are investigated in two different planning domains. Experimental results show that improvements in the recognition process depend on the characteristics of a given planning domain.
展开▼
