Linda is a coordination language using generative communication via tuple spaces, which are global associative memories consisting of bags (or multi-sets) or tuples. The matching of tuples for retrieval has traditionally been undertaken using a simple matching algorithm: performing exact matching or variable substitutions on corresponding elements in a tuple and a tuple template. It is proposed here to generalise the matching process to effectively match according to any function that can be devised as a comparison between tuples. That is, constraint matching, where tuples are matched according to any given constraint. Constraint matching effectively moves processing from the user process to the individual retrieval instruction. In a distributed Linda environment this means a distribution of the processing associated with complex retrievals to the actual data, thus reducing communication and increasing performance. Simulations and emulations of constraint matching in the existing speci...

. In this paper, we present similarity metrics, which provide a seamless integration of cardinal and non-cardinal similarity measures. We show that cardinal and non-cardinal ways of measuring similarity in case bases are simply instances of the definition of similarity metrics, and that a single definition of the maxima of a similarity metric can be applied in both the cardinal and non-cardinal cases, to give intuitively sensible maxima in both cases. We present a number of ways of constructing new similarity metrics from existing similarity metrics. A great strength of the framework is that similarity metrics of different types (e.g. cardinal and non-cardinal) can be combined, and the result will still be a similarity metric (to which the definition of maxima is still applicable). The paper is illustrated throughout by examples from a holiday case base. 1 Introduction Traditionally cases have been retrieved from case bases using weighted similarity measures, in which the similarity o...