1. IF there exists a necessary-ex which has been violated. THEN the model pattern is to be rejected for the data pattern; MODEL-DRIVEN REASONING FOR DIAGNOSIS Jianlai Yan E.P. Schlumberger 50 avenue Jean Jaur~s 92541 Montrouge C6dex, France 2. IF there exists a sufficient-ex which has been verified, THEN the model pattern is to be confirmed for the data pattern; 3. IF (the sum of weights of all the verified verificationex) / (the sum of weights of all the verification-ex) is greater than a verification threshold (e.g. 0.75), THEN the model pattern is to be confirmed for the data pattern. According to the result of verification, knowledge expressed in the IF-CONFIRMED, IF-NONVERIFIED or IF-REJECTED part will be used. We introduce the "model-driven reasoning" as an approach to diagnosing. Lucidness and explicitness of knowledge rep- resentation are emphasized. Good partitioning of diagnosis knowledge in the structure results in the attention focusing of interpretation and gives efficient reasoning power. The approach has been used in the design of a knowledge-based system for identifying paleo-depositional environments in oil prospecting. Advantages The lucidness and explicitness of model-driven reasoning structure allow easy maintenance and alteration. The structure also permits a good partitioning of knowledge. A reasoning stage partition is achieved by separating the knowledge of data pattern formulation, proposition and verification. For verification knowledge a natural partition is realised by model patterns. For example, every time the system checks a proposed model pattern for a data pattern, only the verification knowledge in the model pattern will be considered. Because of this partitioning of knowledge, at any moment, the inference engine pays attention only to the small fraction of the knowledge base which is relevant to the current subject. Efficient reasoning power is obtained. Diagnostic Problem A pattern is a quantitative and structural description of a conceptual object of interest. Simply speaking, the subject matter of resolving a diagnostic problem is to assign data patterns to their corresponding model patterns. Generally, model patterns are domain dependant while data patterns are problem dependant. For some problems, one should formulate at first appropriate data patterns from raw data or other data patterns. Representing data patterns and model patterns on an object substrate is a good approach. Model-Driven Reasoning There are 3 main types of diagnostic knowledge present: (i) knowledge for formulating data patterns, (ii) knowledge for proposing a model pattern for a data pattern, and (iii) knowledge for verifying a model pattern for a data pattern. The knowledge for formulating appropriate data patterns for consideration is very important. They can usually be expressed as production rules. A proposition for a data pattern is defined as a pair of a model pattern and an integer (e.g. between -1000 and 1000) reflecting the likeness (which will influence the order of consideration for propositions) for the proposed data pattern. Proposition knowledge may be represented in the production rule form, like: IF some conditions THEN ($PROPOSE data-pattern model-pattern 700). Every data pattern maintains in it a sorted list of propositions. Most of the knowledge for verifying a model pattern for a data pattern is described in the model pattern by following three expression types: NECESSARY-EX, SUFFICIENTEX and VERIFICATION-EX. The interpretation of the expressions is ensured by 3 rules:

/lp/association-for-computing-machinery/model-driven-reasoning-for-diagnosis-GFQKKUEWr5