Is fundamental to Aititcial Intelligence (AI). Within any Al system intended to exhibit haviou ( be r, in practice, all but the most trivial cases) the pattern matching process inevitably in a combinatorial explosion '. The size of this explosion is determined by the amount of data for a given domain and the number of production rules in use. Currently, many systems spend majority of their time performing the pattem matching task. As a 'reasonable' response time is aspect of intelligent behaviour, clearly, reliable techniques for minimizing the combinatorial aie highly desirable. To date, one of the most successful attempts to achieve this is the Rete algorithm, proposed by Forgy C. L. [For82] for “many patterns/many objects" pattem matching. exploits the redundancy inherent in the matching process, remembering what productio h dam . ns mate elements between cycles and enabling productions to share partial matches of parts of their This affords a con derabl si e gain in eîciency over more straightforward systems which match all rules against all data on successive cycles. combinatorial explosion also affects object-based systems, where knowledge is expressed in terms as o sed ppo to niles. Object-based systems have attracted great interest in the AI community. strong argument for object-based representations is that they offer a natural way to represent and about knowledge. These representations are inherently suited to domains of perception such as qleech recognition and machine vision. In these ?elds the problem of combinatorial explosion is Iorsened by the complex nature of speech and vision. The level of ambiguity present in such real-world III: usually implies that the system must handle and , eventually choose between, many alternative herpretations of the input. The aim of the research described here is to investigate an 'intelligent' matcher for objects. The key 'ssue is to avoid unnecessary computation by exploiting the fact that some pattems share certain components. Using information about previous matches results in the shared parts being matched only once. We believe that by doing so, the eftciency of the matching task can be improved greatly. A imtching algorithm based on these requirements is proposed and an implementation is discussed. The algorithm is intended to recognize general objects described in a relational structure; an object is described in terms of other simpler objects and relations between these objects in such a way that more complex objects can be constructed by combining simpler ones. In this way, knowledge about a plticular task is represented by a hierarchy of structured objects. At the bottom of the hierarchy are primitive objects which cannot be subdivided into smaller components. This abstraction makes the system ideally suited to any domain where objects can be described as above. One potential application of this matcher, the recognition of visual objects such as fonnant shapes in speech s ctro pe grams is evaluated. Another possible application, nainely reading handwritten characters, is also discussed.