Paul Haley needs no introduction. He delivered the first keynote at RulesFest 2011, focusing mostly on the management 0f semantics in systems involving rules for business applications.
Paul opened his keynote drawing a state of the industry at this point in time – walking us through its evolution since the early AI days, from the 5th generation efforts to Watson in Jeopardy. He contrasts the approach taken by Watson to support medical processes (http://www.computerworld.com/s/article/9209899/IBM_s_Watson_could_usher_in_new_era_of_medicine), to what he sees as a more robust semantics and inference based approach illustrated by the Halo project.
Paul states he is much more interesting in knowledge than in rules – going along the same lines what Carole-Ann described as being the key remaining problem in rules management: knowledge acquisition and management. Paul focuses more on structuring and sharing the knowledge than on its acquisition, hence his focus on semantics and logic standards.
His assessment: decision management has knowledge pigeon-holed into black boxes that cannot be re-used elsewhere than where they were originally intended for. There is no BRMS vendor capable of really sharing, and very few actually leverage standards. Paul’s rant does hit a serious problem – one that has been identified for some time and for which numerous standardization and community efforts have been created. The key question for me remains: why have they failed?
He moved on to an assessment of the status of rules engines – disclosing that his work is now based on a Prolog base, which certainly is different. In particular, he asserts (more than once in his slides) that Rete is inadequate to express logic of realistic levels of complexity.
His assessment of the BRMS industry can be distilled down to: BRMSs do not address the management of knowledge, the users are not knowledge managers. And the “knowledge” that ends up captured in those systems are locked into those systems.
Paul then moved on to discuss Natural Logic, of which FOL (First Order Logic) is a formalism. Natural Logic aims at maintaining knowledge in natural language with at least the expressive power of FOL in a manner that remains vendor and tool neutral. Managing natural logic involves working on the expressions – removing ambiguity, simplifying, etc – and resulting in a formalized representation of abstracted high level logic able to express complex problems. Paul is passionate about semantics and higher level knowledge management, and I admire that. He provides deep insight in that currently very complex world (I don’t want to even start counting technologies and standard efforts).
He discussed the current state of the semantics content. Lots of content already available, etc. He gives the example of Watson – which relies as much on probabilistic approaches than on semantics and ontologies – highlighting my key issue with marked-up semantics and built-up ontologies: building and maintaining them with the quality required takes so much effort that probabilistic or frequentist approaches that leverage the data as it is being created and modified are much more efficient and capable to handle change and variability through space and time much better. Google and the rest of the Big Data people take that approach. Remember what Peter Norvig stated as quoted in Carole-Ann’s blog.
This is where I personally take my personal distance with approaches that seek to encode expert knowledge and only rely on that knowledge: regardless of the effort made to have that knowledge as clean as possible, the basis for that knowledge is where the brittleness is, much more than the management of the knowledge. Take as an example driving a car – can we really encode in an exhaustive way the rules of driving, on any terrain, weather, police monitoring situation, local traffic customs, etc? Having lived and driven in 4 continents and more countries I care to count, I can say that ambiguity, local and temporal exceptions, driving habits… all that presents a huge challenge to attempts to codify and manage the knowledge in an effective, realistically reactive way. Google’s self-driving car relies more on experience gained by driving and constantly updated than on pre-built ontologies.
A very interesting and provocative presentation by Paul, as usual!
Learn more about Decision management and Sparkling Logic’s SMARTS™ Data-Powered Decision Manager