Enhancing a production rule engine with predictive models using PMML

In this paper we describe how the Predictive Model Markup Language (PMML) standard enhances the JBoss Drools production rule engine with native support for using predictive models in business rules. The historic debate between symbolic and connectionist approaches to rule/model orchestration provides numerous examples of hybrid systems combining "hard" and "soft" computing techniques to achieve di-erent levels of integration. Rules are often used to decide when and which model to invoke; model outputs, in turn, can be used to evaluate the preconditions of a rule. In a loosely coupled system, the rule engine calls an external component implementing the predictive model, but this has several disadvantages, most notably the need to setup proper communications and reconcile any di-erence in the way the components encode the data. We propose instead, a tightly integrated system where predictive models and rules become part of the same reasoning framework. The models, encoded using the PMML 4 standard, are loaded and processed by a compiler implemented using the rule engine it-self. The PMML document is transformed into a set of facts that de-ne the model, and a series of rules that formalize the model's behavior. In addition, most PMML data processing, validation, and transformation procedures are also implemented using auto-generated rules. Finally, in oder to integrate model inputs and outputs seamlessly in the inference process, we exploit an extension of the Drools engine which adds native support for uncertainty and/or fuzziness.