Widely adopted for more than 20 years in industrial elds, business rules o er the opportunity to non-IT users to de ne decisionmaking policies in a simple and intuitive way. To facilitate their use, systems known as Business Rule Management Systems have been developed, separating the business logic from the application one. While suitable for processing structured and complete data, BRMS face di culties when those are incomplete or uncertain. This study proposes a new approach for the integration of probabilistic reasoning in IBM Operational Decision Manager (ODM), IBMs BRMS, especially through the introduction of a notion of risk, making the compilation phase more complex but increasing the expressiveness of business rules.
Business Rules Management Systems (BRMS), such as IBM Operational
Decision Manager (ODM), have been introduced in the 90's to facilitate editing,
authoring, deploying and executing business policies by domain users, in the
form of conditions/actions rules. Syntactically close to the business language,
these ease the translation of decision-making and business strategies, making
them accessible to users with no programming experience. Besides the rule set,
an object data model describes the di erent objects concerned by the rules, they
are dynamically instantiated in a working memory during the execution of the
program. The activation/execution of rules is managed by inference algorithms
such as RETE [
The PhD thesis presented here builds upon work already carried out within
IBM France Lab and the LIP6. It is based on the hybridization of business
rules with probabilistic graphical models such as Bayesian networks [
A Bayesian network is a compact representation of a joint probability
distribution over a set of random variables. These appear in the form of nodes in
a direct acyclic graph (DAG) where the absence of arcs represent conditional
independences. This type of structure is used as a decision-making tool in many
expert systems and applications [
Probabilistic Relational Models (PRM), on the other hand, are combining
notions from Bayesian networks and from the paradigm of object-oriented
languages [
Several studies including a thesis [
Figure 1 shows how those works t into the ODM toolchain. Several
semantic trees are generated after analyzing and checking the syntax of the user le
describing the rules and the object model (ARL). In addition to producing the
one describing the rules (SemRuleset ) these studies proposed to use the object
model to generate a probabilistic graphical model semantic tree (SemPGM ). The
probabilistic expression of rules are then rewritten to link ODM to an inference
engine (aGrUM1, jSMILE2, ...) using the generated le (PGM le). Once the
rules are rewritten, the models are compiled before being translated into
bytecode in an archive (JAR), facilitating deployment and execution on the target
machines.
1 aGrUM (A GRaphical Universal Modeler) [
A working example used in those studies is that of a fraud detection application, an organization being responsible for managing the reimbursement requests made by its clients according to their nature (type and cost of reimbursement). For this purpose, a set of healthcare professionals are in charge of validating the reimbursement requests made by customers.
Figure 2 shows an example of relation schema for PRM classes of this problem. A HealthcareProfessional class describes healthcare professionals based on several characteristics, including their age, location, gender, and the list of clients they manage (Subscriber), themselves characterized by an age and a list of reimbursements requests (Reimbursement).
Generated during compilation, this model allows to answer a rst set of probabilistic rules. In the example below, we will trigger the action part when we detect a trusted healthcare professional based in Paris who has a client with a high risk of fraud.
{ Business user friendliness: such rules can be di cult to de ne and to understand by a business user, expressing a probability on particular conditions requiring a deep level of knowledge of the probabilistic models used. { Performance: ODM provides the ability for users to de ne di erent types of conditions (for example, using lters, aggregators, and nested conditions). Neither these constructions, more complex, nor their impacts on the performances have been studied.
Incidentally, a new inference algorithm [
We notice that the graphical model built during the compilation of ODM takes into account a probabilized model of the relations between the objects of the working memory but it does not adapt itself to the probabilistic queries de ned in the rules. 2
To address the business user friendliness issue raised above, we need to rede ne the treatment of uncertainty in the expression of rules by replacing the probability thresholds attached to single variables by a notion of acceptable risk on the evaluation of the conditions of the rule as a whole. The action part of a rule will therefore be executed only if the set of conditions is veri ed with a probability greater than the de ned acceptable risk. This will allow our probabilistic rules to be both more complex and intuitive, but it requires a rede nition of the rules compilation phase to redistribute the overall risk to each individual condition.
In the rule below, a rewriting of the previous one, the then part will only be triggered if the probability that conditions c1 and c2 are true is greater than 80%. Using such kind of formulation will ease the use of probabilistic rules by business users.
To achieve this, it is necessary to generate the PRM not only from the object models but also from the set of rules (see Figure 3). All the predicates over attributes of a probabilistic object, such as hp.risk==low, will be added in our graphical model in an extension of the concerned class (red classes in Figure 3). This requires studying the di erent forms that conditions can take (aggregations, lters, nesting, ...) and their predicates. Hence, a feasibility study must be carried out to de ne the theoretical and practical limits of such a method. The rules will then be rewritten to take into account our new probabilistic queries (on the probability attribute of the detectInvoiceAnomaly class, for example). Such manipulations could, however, change the e ciency of the pattern matching performed by ODM for the selection of rules to be performed, therefore their impact should be studied.
On the PRM side, the dynamic instantiation of classes representing rules (in green in Figure 3) as well as the simpli cation of the model generated during compilation will be at the center of the project (by taking into account that some of the data sources will be deterministic). In our example the location of a healthcare professional is considered certain, thus we don't need to include it in our new PRM.
Due to the highly deterministic aspect of the user models, we will assume that the instantiation of our PRMs will yield a large number of small disjoint subgraphs (one per health professional, for example), which would require working on a new inference algorithm taking advantage of this structural redundancy.
This development work is based on what has been done in the Bayesian Insight Service (BIS) plugin developed as part of the previous thesis. The gure 4 illustrates how our new module, PRIME (Probabilistic Reasoning Insight ModulE ), ts into the ODM compilation chain. It intervenes directly during the process of rewriting the semantic tree describing the rules (SemRuleset ) but, unlike BIS, extends the de nition and optimization of the graphical model from the rules before rewriting them (via the PRM enhancement process).
Once the rules are rewritten, the graphical model is serialized in order to be usable by a probabilistic engine in parallel with ODM.
As part of our work, ODM uses aGrUM for probabilistic calculations and manipulation of PRMs. The complexity of the compilation, as well as performance problems during the execution, might require working on an extension of this library (advanced aggregations, functional nodes, parametric nodes, ...)
This work will provide an extension of the ODM business rule de nition language to probabilistic concepts. Syntactically and semantically close to natural language, it will allow non-computer scientists to de ne probabilistic business rules in an non-ambiguous and intuitive way.
A study on the use of temporal models via an extension of the PRM
modeling capabilities in O3PRM3 could be done in a second step, which would allow
a similar reasoning but in the context of the treatment of complex events
(Complex Event Processing, CEP [
Finally, one of the advantages of the expressiveness of PRM is that it is
possible to take into account the structural uncertainty of a model [