Paul Haley has posted a clear and easy-to-read look at the symbiotic nature of analytics and business rules in the field of “Decision Management”. Furthermore, he goes on to describe a run-time merger of these concepts under the term “Adaptive Decision Management” [*1] or ADM.
Some thoughts on ADM from a CEP perspective:
- Decision performance monitoring is like BAM (or maybe “metaBAM”). Rule firings are events. CEP might be useful to correlate rule use versus data being processed, just as CEP is useful in monitoring BPM, SOA and so forth.
.
- Running continuous statistical monitoring of rules and rulesets requires some stateful handling of the real-time models [*2]. But a stateful rule engine (like that used in rule-driven CEP) should be able to easily handle that.
.
- You probably want to cross-correlate “decision performance” statistics (and the real-time equivalents of certain analytic models) across different decision services and decision types. Meaning across different event types. Which means CEP again.
.
- Proponents of conventional SOA decision services and analytics tools will probably point to (parts of) PMML as the bridge between analytics tools and rule engines. This is very true, but doesn’t avoid the fact that analytics on historical events is, well, historical. In other words it is already out of date, which might (or might not) matter [*3]. So at the very least, doing real-time/”operational” analytics can complement conventional analytics [*4].
.
ADM indicates another step in the convergence of BI, BAM and decision management, and could well describe some CEP use cases today.
.
Notes
[1] A quick Google search shows Paul to have successfully found an almost unused Three Letter Acronym (although of course the abbreviation has prior art).
[2] Possibly you could treat analytical / statistical results like other data and store them in a data service (/database). But that would require administration, and might be more difficult to handle across distributed rule services. Of course these aren’t unsurmountable – for example you could use some flexible (i.e. schema-free) data repository like an RDF store…
[3] For example, in fraud, finding a new fraud pattern next week doesn’t help this week’s fraud victims.
[4] At least one conventional Data Warehouse company seems to agree with this!
