Taux d'encadrement : 50%

Thèse CIFRE IBM

Dates : 2017-

Probabilistic Rules Optimized COmPilation

PROCOP - Probabilistic Rules Optimized COmPilation- is a PhD offering to investigate advanced compilation of probabilistic rules in complement of effective, and scalable ways of enhancing business-rule based decision- making with the capacity to take uncertainty into account in an industrial context. The maturity of both Business Rules and Bayesian Reasoning technologies has reached a stage where they can both benefit from each other.

As the world goes more complex, the decision-making of companies needs to be more refined and flexible. Therefore, the concept of “Business Rule Management System” (BRMS) was introduced over a decade ago to facilitate authoring, checking, deploying and executing the business policy of companies as conceived by their business, as opposed to technical, staff in the form of condition/action business rules. On the other hand, Bayesian Networks (Bns) – also called Probabilistic Graphical Models – were proposed in the late 80s for modeling uncertain knowledge and reasoning with incomplete data. They were initially used as tools for probabilistic reasoning in early expert systems [5]. Many improvements and dedicated algorithms have been proposed since then [6]. Particularly new languages which enrich and structure the models by using first-order logic or Object-Oriented extensions (O3PRM, [2]). Today's knowledge-based software systems have to inter- operate. Moreover, the emergence of the "Big Data" processing emphasizes the importance of analytics and probabilistic modeling of data. Hence, adapting the Business Rules to uncertain reasoning will become essential.

For this reason, the PhD aims at building a bridge between two separate « worlds » by introducing the notion of Probabilistic Production Rules. In effect, we propose to specify and implement valid extensions of actual BRMSs to enable them to take into account probabilistic information. As well, it will investigate valid extensions of BNs dealing with rules, implementing these extensions and validating them on real-world decision-making applications. It continues a previous study named URBS PhD undertaken at IBM proposing a first approach for implementing Probabilistic Rules. PROCOP main objective is to improve the rule compilation proposed for URBS by extending the supported uncertain pattern matching expressions. This task will require to introduce new operators and instructions for the definition of the underlying graphical models. Meanwhile, the rule compilation should be highly compliant with the IBM cognitive platform, especially with SPSS in order to benefit from its predictive models.

The first task will deal with theoretical work, with the Lip6 lab from Jussieu University having experience and expertise in both the theory and practice of graphical models. The challenge in this area is to integrate complex knowledge and decision representation into Bayesian reasoning. Indeed, there are two main questions: (1) how to extend standard model and rule language for supporting probabilistic reasoning by using the concept of risk of occurrence, and more generally, (2) how to perform efficient reasoning over such knowledge using production rules. Additional questions may arise about how to build/induce effectively probabilistic production rules from data.

The second task will focus on the interoperability of the BRMS IBM ODM with probabilistic and predictive engines of existing frameworks (SPSS, Agrum). The goal is to process by iterative experimentation and validation in real use cases to be defined during the PhD.

Ces informations sont issues du site HAL.