Innovative algorithm and methodology for expert strategies on-line continuous extraction, control and updating
The problem of extracting and updating a rule basis is crucial in Machine Learning. The followed methodology takes into account a cognitive model of the operator. According to this model, the operator uses a limited number of rules, each of them involving a few number of parameters. As a mathematical consequence, the current set of decision examples (also called control situations) can be viewed as paved by small dimensional “cylinders”. This constraint structuring principle of the rule space is used to: - From an initial set of control situations, extract a set of rules representing the related expert decision rules in controlling the industrial process. This has been achieved in tuning decision trees algorithms so as to fulfil the basic cognitive hypotheses mentioned here above. - Manage the cylinders and rules so as to emphasize in real-time possible inconsistent inputs thanks to the notion of “check as you decide”(implemented through the decoration function, that use a colour code related to decision correctness) - Locally update the rules set by elementary operations on cylinders. This updating appears to be efficient to solve local inconsistencies, despite the fact that the general problem of managing inconsistencies is NP-hard. This methodology led to UML specifications that allowed the implementation of the algorithm and the design of a user –friendly man-machine interface. These results may be considered in a more generic way as representative of a high level methodology consisting in relying on human intervention and role in process control, while setting a human centered expertise control loop thanks to a continuous experience synthesis and feedback. The overall scientific result of the project are a concrete contribution to the formalization, analysis and implementation of Anthropocentric Production Systems.