1. Reasoning at a point within a [business] process
2. Reasoning about events that occur over time.
3. Reasoning about a [business] process (as in deciding what comes next)
4. Reasoning about and across different states (as in planning)

Enterprise decision management (EDM) addresses the first.  Complex event processing (CEP) is concerned with the second.  In theory, EDM could address the third but it does not in practice.  This third item includes  the issue of governing and defining workflow or event-driven business processes rather than point decisions within such business processes. 

Business applications of rules have not advanced to include the fourth item.  That is to say, business has yet to significantly leverage reasoning or problem solving techniques that are common in artificial intelligence.  For example, artificially intelligent question and answer systems, which are being developed for  the semantic web,  can do more than retrieve data – they perform inference.  Commercial database and business intelligence queries are typically much less intelligent, which presents a number of opportunities that I don’t want to go into here but would happy to discuss with interested parties.  The point here is that business does not use reasoning much at all, let alone to search across the potential ramifications of alternative decisions or courses of action before making or taking one.  Think of playing chess or a soccer-playing robot planning how to advance the ball on goal.  Why shouldn’t business strategies or tactical business decisions benefit from a little simulated look-ahead along with a lot of inference and evaluation?

Even though I have recently become more interested in the fourth of these areas, I expect the audience at the business rules forum to be most interested in the first two points above.  There will also be some who have enough experience with complex business processes, which are common in larger enterprises.  These folks will be interested in the third item.  Only the most advanced applications, such as in biochemical process planning, will be interested in the fourth.  I don’t expect many of them to attend!

The notion of enterprise decision management (EDM) is focused on point decision making within a business process.  For enterprises that are concerned with governing business processes, a model of the process itself must be available to the business rules that govern its operation.  I’ve written elsewhere about the need for an ontology of events and processes in order to effectively integrate business process management (BPM) with business rules.  Here, and in the workshop, I intend to get a little more specific about the requirements, what is lacking in current standards and offerings, and what we’re trying to do about it.

As I’ve written previously, the distinction between business process management (BPM) and CEP is not well principled but arises from somewhat arbitrary, historically distinct emphases on technology and market segmentation.  Any modern business process system must handle events and discussing events processing without considering processes is a limiting perspective.

Most people would intuitively agree that events trigger business processes.  For example, a business receiving a payment or a letter from a customer or vendor is an event that triggers the process of crediting and depositing that payment or considering and responding to that letter  Hopefully, we are moving beyond academic arguments about the distinction.  (I am ignoring here the algorithmic applications of streaming event processing as in trading in the capital markets.)

So, in current BPM, which should include CEP capabilities, we want decision management that is less ignorant about where it is in a business process and that is aware of events that trigger processing.  That is, we want policies that talk about the state of a business process and the occurrence of events.  The surprising thing is that current business rules management systems (BRMS) and related standards are of no help.  Tools from the leading business rule management system (BRMS) vendors, including Oracle and IBM / Ilog, have no intrinsic understanding of processes, events, or time.  And, outside their integrated BRMS, tools from BPM vendors don’t let us “talk” about anything.  They help us structure flows and code, but they rely on integrated BRMS to manage rules.  The rules include the policies, in which the business “talks”.  The BRMS is where English sentences (or something that increasing looks like English) are managed as the enterprise repository of policy. 

Note that there is nothing special about English.  It’s just easier to read than “natural language”.  Another limitation of current policy management systems is their lack of language independence, which requires automatic translation, which is much simpler if the sentences are unambiguously interpreted with logical rigor, but I digress from the point of this missive…

Some examples will help here.  Suppose we check the credit of an applicant at various points within the collection of processes that constitute how our enterprise conducts business.  We might have policies that are concerned with how we consider or act based on credit information in originating a loan (or policy) versus in renewing  or re-pricing one.  In effect, our policies want to talk not just about evaluating credit or pricing risk, but to do so in the context of a larger business process.  To be more specific, business policies that sound like, “if evaluating credit in the course of pricing a renewal…” or “if evaluating credit while considering a new policy…” are quite natural.  These statements define or govern the business process.  They also talk about where the decision is being made within a business process.  

• BRMS need to understand the context of the business process in order to make context-sensitive decisions.
• BPM needs to tell the BRMS what it is going on from the top-down for the BRMS to understand the context of a decision.

So, we need the BRMS to be told things like:

• I am considering a new applicant.
• I am considering the renewal of a contract.

In these statements, the pronoun “I” is the overall enterprise system contemplating its own actions.  If you find that awkward, just substitute “while”  to obtain what you might “say” in a business policy.  Ideally, the language your policy management system would not be overly stilted but would understand any of:

• while considering…
• during consideration of …
• if considering…
• if an … is being considered for…

and so on.  Otherwise, users will find authoring such statements cumbersome.  Reading and understanding English, even if it is a bit stilted, is easy for people.  We’re built to communicate, after all.

Now consider what you would want to say if you were writing policies that involved events.  In this case, the event has already occurred, such as “we received a letter from a customer…”.  You do not want to say, “if I am receiving a letter from a customer…” (which could only be true for an instant that passes quickly unless it was stated as “… I will be receiving…”).  But if we can only refer to events in past tense, how do we talk about a current event that needs to be handled versus another event that we have already handled?

Many business to consumer (B2C) applications, such as pharmacy benefits have this problem, for example.  To a pharmacy benefit manager (PBM), like Medco or Express Scripts, the swiping of an insurance card at a retail pharmacy is an event to be processed.  Any individual beneficiary has a history of such requests.  We can try to model the current one as a request and the prior ones as transactions, but this becomes awkward for less formal or technical people who want to talk about how many requests someone has submitted over a period of time, for example.  The truth is that there is a history of requests per beneficiary and technical limitations should not obscure this fact.  We should be able to distinguish the current request from prior requests, as in the following:

• a request that has not been processed is current or pending
• if processing a request…
• if a request for … is being considered…

Note that “request” is a deverbal noun, which is to say that the root form is the verb (in this case “to request”).  A request is a reference to an act of requesting that may be in any tense.  The sentences above reflect this in the use of an additional verb that carries the tense.  Of course, this is all completely natural since every sentence has a verb.

The most dangerous expression might be:

• if … requests …

and yet this is the form that almost all BRMS would handle today!  This is dangerous because it is too ambiguous about when the request occurred.  It would be better to say:

• when … requests …

provided that the system understands that, unlike “if”, “when” involves time, but even “when” is less than ideal since an event has always occurred in the past by the time it is processed.  On the other hand, we might define when a request occurs as in:

• a request occurs from the time it is received until it receives a final response.

This assumes that “when” combined with a verb in present perfect tense means during the period of time in which the process referenced by the verb continues.  And this is an important point:

• An event can be an occurrence of a process.
• An event may have a duration.
• “When” may refer to an interval of time.

Events are not necessarily processes, but may refer to instantaneous points in time, such as in the following:

• When a request is received…
• When the processing of a request begins…

but these uses of “when” refer to a point in time before any action can be taken in response to the event, therefore the sentences should only conclude with statements of implied, necessary, or modal logic and should not include any statement of action. Of course, a competent BPM/CEP/BRMS would understand all this and advise the author of a policy that suggests taking action in the past.

As we proceed through these examples our intuition should be building the understanding of the first three points made above.  Processes and events and reasoning about or over time are completely intertwined in nature and separating them between BPM and CEP and BRMS systems is completely artificial and hopelessly limiting.

So what is the solution?  I suggest it is a knowledge management  system that understands the following:

• 1. Policies that use tense.
• 2. Policies that refer to events using deverbal nouns.
• 3. Policies that refer to occurrences of processes as events.
• 4. Policies that refer to potential action using future tense, possibly by way of modals.
• 5. Policies that refer to occurrences of processes using verbs such as “begin”, “end”, “start”, etc.
• 6. Policies that refer to occurrences of process using words like “during”, “while” and “when”
• 7. Policies that refer to events using prepositions like “by”, “before”, “after”, and “when”

The natural language technology to parse such sentences is widely available using many approaches.  I am happy to discuss that with interested parties.  The second step that needs to be addressed is transforming the logical interpretation of such sentences derived from the natural language system into the underlying execution architecture, which includes a process engine and a rules engine that must be appropriately integrated.  That integration involves the informing of the BRMS about the state of the business process and the actions that may be taken which may be expressed as processes in the BPMS.  I’ve written elsewhere about this in more detail and am also happy to discuss it in more detail with BPM or CEP practitioners, product managers and architects.

Understanding events and occurrences of processes as events adds a great deal of power to policy management.  It allows statements of policy to reference and consider the context of business processes.  It allows statements of policy to reference and consider how to handle events in the context of business processes.  And, if it is done with adequate natural language understanding, it accomplishes this integration of BPM and CEP within a single policy management system.

Although I had hoped to cover the fourth point made first above here, I now prefer to conclude with a brief discussion about reasoning over time.  I will strive to cover reasoning about potential states of a process another day.  It is interesting but rigorous material that requires (in my opinion) architectural support that is lacking from current rules engines, whether production rule or logically based, even if the situational or event calculi are good formalisms.

Reasoning over time is pervasive in CEP.  In the pharmacy benefits management domain, for example, coverage is commonly limited based on the history of transactions.  For example, a policy might limit the amount of refills over a period of time.  This involves aggregation over a number of events, each of which is the result of handling a prior request.

Very few knowledge or policy management systems understand that transactions are processes, occurrences of which can be viewed as events.  For example, is “order” an noun or a verb in your enterprise applications?  Our technology has biased us to thinking about objects, which drives our modeling towards nouns and away from verbs.  Our technology biases us against modeling events and processes well!  And it shows up, insidiously sapping productivity and accessibility.

The lack of ontology of process and event in current BRMS not only precludes the kind of integrated BPM and CEP I am discussing here, it also limits the ability of current BRMS to automate policies that consider what has happened in making decisions in the present.  For example, a statement like:

• if a medicine has a maximum therapeutic dosage over a period that is less than the total dosage of that medicine requested by a member over the same period then…

is beyond the capabilities of current offerings.  Authority understood some grammar about time but did not understand that events, such as a request, occurred in any deep sense.  So it could automate a sentence like:

• if the total dosage of a medicine requested by a member on a date within the last 90 days exceeds the maximum quarterly therapeutic quarterly dosage for the medicine then…

but understanding why it understands one sentence and not the other is too much for many authors to tolerate, let  alone understand.  The essential reason is that we sold the company before extending Authority’s ontology to include events and revising its parser to understand that both verbs and their deverbal nouns referred to events (including occurences of processes).

The bottom line here is that a quantum leap in natural language processing of business rules is needed.  Fortunately, this is not a quantum leap for natural language processing itself.  It is well-established that sentences are parsed into representations of events in which noun phrases play semantic roles, such as the following:

• agent or counteragent
• donor or beneficiary
• patient or “experiencer”
• causer
• instrument
• locative or time
• source or destination

where quite a few prepositions relate to more refined aspects of time and location, such as at, on, during, by, before, after, in and so on.   The critical thing for processes and events is that they occur in time.

Realizing this quantum leap in policy management and knowledge automation is really pretty simple.  Take an approach such as Authority and extend its core, upper ontology with the semantic roles and the concepts of events and processes.  Then extend its relation-centric parsing with even-centric parsing (both are needed).  A few more steps, notably handling metonymy, and the next generation of knowledge management and automation that provides the integrated understanding of time, events, and processes discussed here becomes a reality.

That’s what we’re patiently working towards.  And we’re doing it in as engine-independent a manner as practical so that we can leverage standards like RIF and SBVR.  It’s all about the knowledge.

Finally, we are looking for collaborators who would like to learn more or help, and, perhaps, get involved in leveraging the solution or its underlying technology.

David Karger of MIT’s AI lab challenged the panel from the audience by asserting that the data people posted on the web is much more important than any ontology that might define what that data means.  This set off a bit of a firestorm.  There was consensus that data itself is critically important, if not central.  For the most part, panelists were aghast at the notion that spreadsheets of data would be useless to computers unless the meaning of its headings, for example, were related to concepts defined by reference to ontology those computers understood. 

With respectful deference, the panel and audience yielded.  Sir Tim Berners-Lee took the floor.

The issue of semantics briefly faded from the discussion.  Utility seemed to be the crux of the matter.  Sir Tim illustrated how even the smallest bit of semantics (i.e., meaning by reference) added to a spreadsheet allowed others to quickly and incrementally, almost continuously add value to published data.  He did this by way of example, discussing mash-ups of bicycle accident, traffic, and map data by different people over the course of a day or so after someone first published the bicycle accident data.  Most interesting to me, however, was his concluding point: simply identifying what something is by reference to an existing semantic web concept makes data much more immediately consumable, useful and valuable.  For example, simply identifying that a column is a longitude using RDF adds a lot of value.   Yes, I realize mash-ups are old, and linked data is well known, but his point was delivered in such a straightforward and compelling manner that the argument simply passed.  He put  a lot of experts eyes back on the ball.

Sir Tim reiterated this point an hour or so later at a meet-up of meet-ups.  His consistent, critical point was that lightweight use of ontology realizes a great deal of value.  Simply using RDF to anchor the semantics of linked data, without substantial ontology development, makes data much more useful to humans.  And with regard to enterprises, he discussed how eleven (yes, he said 11) different concepts are enough to address most of the semantic needs of various enterprises (i.e., their relational database models).  Of course, they will receive more value from reusing existing ontology, extended according to their needs.  His point was that the small step of simply linking data with open web semantics using only the most widely adopted RDF identifiers is a huge step forward for the semantic web and the benefits its technologies can bring to individuals and enterprises.  Full-fledged ontology is important for deeper functionality and long-term visions, but simply using concepts from existing ontology can be a huge step forward.

By no means is this a book about SBVR.  Rather, it is about the core and critical concepts necessary to understand how  to define an enterprise repository of declarative knowledge covering business.  The unfortunately reality is that the work necessarily suffers from legacy concepts and perspectives, but this is a statement of fact and practicality rather than a negative assessment.  Essentially, my problem with much of the increasingly mainstream marketing and practicing of business rules is precisely the term “rule”.  Nonetheless, since the mainstream audience is pre-occupied with the word “rules” rather than the word “knowledge”, Ron’s book is – again – right on the money for practitioners who want to achieve the benefits of separating models and knowledge from implementations in order to increase the agility of systems and more closely align IT with business needs (all of which he discusses in Chapters 2 and 3).

This edition clearly benefits from the progress in the field and reflects current enterprise objectives that have progressed into the mainstream since Ron began practicing and writing many years ago.  Veterans will quickly recognize and benefit from the simple clarity that Ron brings to the subject.

The critical topics that Ron addresses are the core concepts underlying modeling and linguistics that allow non-technical business analysts (even, perhaps, subject matter experts) to capture unambiguous statements of business definitions, requirements, policies and other “rules” in English sentences that are suitable for presentation to and verification by stakeholders.  In addition, in the event knowledge-level standards like SBVR evolve and are adopted by technology (e.g., business rule) vendors, managing such statements will directly impact the operational behavior of business processes, especially where they involve automated decisions, policy enforcement and other forms of governance requiring enforcement.

Although Ron and I have come to a shared view on knowledge capture and management, we have followed different paths.  Ron’s concern is primarily capture.  Mine extends to and necessarily emphasizes implementation and execution.  Nonetheless, our perspectives on modeling and capture are extremely aligned.  Ron’s writing on what he calls “terms” and “wordings” are right on the money.

We have independently arrived at a point where we believe that linguistic expression of business knowledge is much more important (and valuable) than rule-based expressions suitable for execution by rule engines.  In effect, Ron argues and I agree that the knowledge is much more important, durable, and valuable than its executable form or expression.  As a result, I could not more emphatically introduce Ron’s coverage of how to express models and logic in English.

By the end of Chapter 1, Ron defines what he calls a vocabulary and why  managing  vocabulary is a requirement for managing business knowledge (including definitions, requirements, and policies).   Ron’s writing on nouns and verbs and how they form the backbone of expression for business knowledge is seminal and accessible to novice and expert practitioner alike.

As more or an artificial intelligence practitioner, I might quibble with the omissions of mass nouns, such as money or time, from his discussion, but these topics will not be missed by anyone but the most forward-looking strategist.   Similarly, the limitation of current standards to wordings involving verbs eliminates coverage of adjectival phrases and results in too many wordings involving the verb “has”, at least in my opinion.  For example, wordings like “a person has an age” seem arcane, perhaps technical, to me.  Nonetheless, even though my work supports such “phrasings”, most of its users still use “has” phrasings!  Perhaps this is because I have not taken the time and expended the effort to clarify the concepts that Ron covers so well.  In short, Ron’s writing is right on target for practitioners given the current state of the art (and standards).  And I can think of no other expert or author who has reduced current practice to such an accessible form.

Chapter 4 clearly explains and motivates the use of sentences as the independent units of business knowledge.  Ron does a good job of introducing the notion of roles as the components linked together within predicates using verbs.  This is fairly abstract material, so I don’t blame him for avoiding too much depth on the semantics of roles.  Fortunately, other materials, such as the SBVR standard itself, cover these details in greater depth.  So again, the content seems right on for someone entering the field and focused on linguistic modeling, capture and management rather than more formal “semantics”. 

Ron clearly separates statements defining and relating concepts from requirements.  Not only does he present the material linguistically, he reinforces it with clear graphical presentations of the corresponding models.   Personally, I prefer to mix certain requirements with definitions, as in “a person has exactly one mother”,  but this would complicate the graphical presentations.  He also addresses common linguistics, such as passive voice and using verbs (i.e., participles) as adjectives, which I have found to be important concepts that must be brought to practitioners’ attention.  Otherwise, the models become unnecessarily complex or the resulting sentences seem cumbersome, if not stilted.

In places the formality of SBVR shows through.  For example, “a person must not lease a vehicle the person owns”.  This reflects linguistic limitations of available tools.  This phrasing reflects “necessity” within the underlying logical formalism of SBVR.  In addition, the reiteration of “the person” begs for natural language processing that would understand pronouns, such as “(s)he”.  Having said that, this is clearly a higher level statement than can currently be understood and implement by most business rule management systems (BRMS), however.  As such, it demonstrates the power of Ron’s approach and motivates (given most commercial offerings) the separation of capture and management of business knowledge from its expression in business process management systems (BPMS) or BRMS.

Having encountered the notion of logical necessity in the preceding example, it is worth pointing to Ron’s discussion in Chapter 10.  Externalizing business knowledge from systems requires certain architectures for information systems.  For example, the prohibition of a person leasing a vehicle that (s)he owns requires some action not stated here.  In effect,  a violation of requirements must be anticipated in the runtime architecture.  There are various approaches to this, including runtime exceptions.   Ron demonstrates this approach in Chapter 10 and, if the system incorporates workflow for resolution, he advocates using the end-user accessible form of business knowledge for explanation (in his discussion of guidance within Chapter 2).  More advanced approaches, such as meta-reasoning, and detailed architectural approaches are appropriately left out of this work on capture and management.

Interestingly, in Chapter 7, Ron suggests that business rules should not be expressed within an if-then syntax.  Obviously, this clearly separates Ron’s methodology from the production rule focus of most commercial BRMS.  It also exposes a gap between SBVR and operational systems.  Ron discusses behavioral rules within the chapter but the gap between requirement and behavior remains unfilled.  This is my most significant concern for the logical approach to knowledge management.   Without a framework by which declarative knowledge can bridge to imperative action, the logical approach falls short of operational relevance.  This is reflected, in my practical viewpoint, by the lack of operational deployment of SBVR. 

The principal  challenge remaining for SBVR is to cross the gap from expression into operation.  Aside from linguistic limitations, this is my primary concern for the formal logic approach to knowledge capture and management.   The ideas Ron covers so well are necessary for long-term success in automating managed knowledge, but they are not enough.   Unfortunately, most of the business rule vendors do not adequately address the necessary capabilities that Ron’s methodology and toolsets handle.  This leaves us with a continuing need for business rule engineers to bridge the gap, manually.

I started to use Radar Networks’ Twine at the invitation of CEO Nova Spivak after writing this earlier this year (also see this). I enjoyed it for a while, especially because a lot of technology folks were hooking up with each other, especially the semantic web community, on Twine. But I found it  tedious to work through beta issues and to be bothered with recommendations or news about who was saying or bookmarking things about what. (I should have turned off the emails sooner!)

I was disappointed that Twine was taking an apparently folksonomic approach to tagging. It was as if Radar Networks was riding semantic web buzz without really embracing it openly or sharing the momentum that the invite-only community was investing in.  That may not sound fair – I believe that there are semantics in the back room, but that’s how it felt and it’s still the way it looks.  But probably the worst part is the process that you have to go through to add a bookmark – which is the whole point, of course!  (I ultimately sacrificed popup blockers, but the process still seems laborious compared to other alternatives.)

I stumbled across Zigtag almost accidentally while working for a VC firm with a portfolio of semantic startups. What I like most about Zigtag is that they make it obvious that they are building an ontology of tags and encourage users to select semantic tags (i.e., concepts) rather than folksonomic “words”.  They also provide tools for managing tags that allow you to move smoothly and incrementally from a folksonomic to a more semantic approach.

The key to the semantic approach for Zigtag is that shared tags are just that – they are more precise than strings.  They are not only words – they have definitions.

Unfortunately, like Twine, Zigtag’s ontological model remains hidden.

My initial experience with Zigtag resulted in immediate jubilation.  The Firefox plug-in works for me.  It lets me type in tags with nice completion and recommendations from the tags that others have defined.  Within 15 minutes I was writing to compliment Zigtag on a practical, elegant approach to the semantic bookmarking problem.  I liked it much better than Twine right off the bat – and despite its book-market, I like Twine a lot!  Within a few minutes I had an email from their founder, Reg Cheramy.  An hour later we were talking.  We talked about his early meeting with Michael Arrington, how his work compares to bulletin board or discussion forum emphasis in Twine , how he facilitates semantic tagging given a very large ontology and vocabulary, and so on.

Whether Reg took my advice to emphasize groups more or was already headed in that direction is unclear, but Zigtag now has group functionality that seems as good as (and in some ways better) than Twine’s.  If you go to Zigtag the web site, you can find groups to join, but unlike Twine’s web site, Zigtag does not recommend groups for you based on your interests.  I’m not sure this is a problem, though.  Recommendations can be distracting.  Nonetheless, if people want recommendations for more than content, it would be a simple step for Zigtag given the fact that they already recommend content that others have bookmarked.

I’m not too concerned with recommendations, even of content, so I cannot comment on Zigtag versus Twine on that front.  Generally, there is plenty of RSS and recommendation noise to go around.  I prefer the linked approach to finding information rather than searching and I don’t expect recommendations to become excellent in the near term.  For more on this, you might want to check out the recent news  about Vulcan’s EVRI investment at Webware or ReadWriteWeb.

I like to use Zigtag from the sidebar in Firefox.  Actually, I owe Reg additional thanks for, in effect, causing me to abandon Internet Explorer for Firefox.     I use it primarily to organize my bookmarks semantically and across machines.  For those that want to do the same, you might also be interested in Mitch Kapor’s Foxmarks.

I’m fine with finding groups on my own and I like seeing people and what they tend to tag, too.  Now that I know they are available on the web site, though, I want them in the sidebar.   The fact that they are indirect on the web site, not presented in the sidebar, and not proactively recommended probably explains why there are relatively few (especially compared to Twine).  It would be nice, for example,  to see groups and people organized along with bookmarks according to how heavily they use tags as I pivot through various facets.

So, on a feature basis, I like Zigtag more than Twine for two primary reasons:

1. Zigtag’s Firefox plug-in is a great user interface while Twine’s book-market is awkward in every sense that matters to me.
2. Zigtag emphasizes and leverages shared tagging of tags that have clearly documented interpretations  Twine is too folksonomic.

The picture shown in this post shows that Zigtag already “knows” a lot about semantics.  Part of the reason is that they must have a roomful of people watching for  tags that people enter that are not defined.  Quite a few of the tags I’ve added have become defined within hours (sometimes minutes) of when I enter them.  We’ll see how this scales up, but I like it – a lot.

The key question for both these sites is:

Are you going to share your ontology?  If not, why not?  If so, when or why not now?

Note that I am not suggesting they should. But if they have a reason not to, it would be nice to understand that.

It also would be nice to know whether the effort I expend on either site will be lost if they are acquired or I want to switch.  That’s how it looks at Twine today.

Zigtag exports my bookmarks.  I can get them from or over to Delicious, no problem.  But I want their semantics, too.  I would really appreciate preservation of the text, preferably the semantics of my tags.  Perhaps if my bookmarks were simply output as an OWL referencing their ontology?  At least then I could move without losing the effort that I have put into them, whether folksonomic or semantic.  I also want to know if their ontology is are any good and, if so, I’d appreciate export to OWL so that I could use bookmarks for other purposes that interest me.

The background issue of data portability, for bookmarks, social networks, and other personal profile data is huge.

If I had OWL export and an open ontology, I would be less worried about my investment in Zigtag or Twine.  Consider Techcrunch’s recent comments:

Zigtag’s biggest obstacle is the slew of other social bookmarking sites already available (…). The semantic tagging feature is fairly unique, but its appeal is still untested, especially against automated semantic taggers like Twine. Frankly, a lot of people are just going to stick with the simple but effective Delicious interface.

It’s hard to argue with the first sentence, but the second seems harsh.  Twine is getting credit that it may not deserve.  Also, Zigtag recommends tags, too.  But the third sentence is a problem for Zigtag as well as Twine, although the latter benefits from superior PR.

Another question, of course, is how Zigtag and Twine will fare once they try to make money.  Radar Networks has stated that Twine will start running ads by the end of the year.  Zigtag has made no public announcements.  Delicious selectively advertises (e.g., on search pages), perhaps to feed intelligence to Yahoo’s advertising network.  The advertisements are so selective that the value of other book-marking sites may be limited to the intelligence that they provide to established advertising networks.  If so, this will hold down valuations and slow innovation.  We’ll see, but obviously, I hope not..

The FICO score is not a probability. 

Fair Isaac people describe the FICO score as a ranking of creditworthiness.  And banks rely on the FICO score for pricing and qualification for mortgages.  The ratio of the loan to value is also critical, but for any two applicants seeking a loan with the same LTV, the one with the better FICO score is more likely to qualify and receive the better price.

Ideally, a bank’s pricing and qualification criteria would accurately reflect the likelihood of default.  The mortgage crisis demonstrates that their assessment, expressed with the FICO score, was wrong.  Their probabilities were off.

• Was the FICO score a useful metric of creditworthiness before the crisis but not during?
• Is the FICO score a reliable metric going forward?

In these mid-crunch days, Fair Isaac is reminding its customers that the FICO score is a ranking not a probability.  The underlying point they seek to make is that the relationship between the FICO score and the probability of default is more complex and dynamic than their banking customers understood last year.  (Another post on predictive analytics also discussed stationarity.)

It’s the probability that matters, not the score!

In his keynote, Ian Ayres also focused on the inadequacy of scores.

He was explicit that bankers need the probability of default and, further, that they need to know how reliable such probabilities are.  As an example, he cited polls where one candidate is leading by 6 points within a margin of error of 3 points as almost meaningless.  More meaningful would be the probability that the leading candidate will win.  Even better would be an estimate of the probability of default along with an assessment of the reliability or accuracy of that probability.

Bankers increasingly understand that the FICO score is not the probability of default that they need when originating and underwriting credit.  As a result, bankers increasingly understand that there is no adequate external source of the probabilities they need in order to optimize their portfolio performance. 

This realization has several ramifications:

• A market opportunity for predictive analytics in credit has opened on Fair Isaac’s turf.
• Scorecards have lost much of their “solutions” luster, becoming just another technique.

But several things also became clear as I talked with numerous practitioners last week.  Fair Isaac doesn’t have much competition.  In fact, it is shocking how little competition they have in such a large and lucrative market. 

What decisioning market?

Although there is a market opportunity for more rigorous decisioning solutions, there is no significant challenger to Fair Isaac.  I expected to hear more about the Experian Group, but the only direct competitor identified by more than one person was Austin Logistics.  Several people indicated that they were using statistical tools directly, especially SAS, and Fair Isaac itself is placing a great deal of emphasis on its own predictive analytic tools, especially Model Builder.

Note that Vantage Score is really having an impact on Fair Isaac scoring revenues, as reflected in their most recent earnings call transcript.  So there is more competition than may seem apparent to the audience that attends Interact.

Another chasm to cross

Generally speaking, this market needs the benefits of broader machine learning techniques, such as statistics, and a more rigorous understanding and emphasis on probabilities.  The audience, however, is not technically sophisticated enough to become aggressive adopters, despite recent harsh lessons.  In the same panel, every banker in turn solicited risk analysts to join their organizations, headquartered in Asia-Pacific, London, Canada, and on the west coast.  They also agreed that it is easier to learn finance than analytics.

The market for analytics is crowded with sophisticated tools and intellectually demanding techniques that are simply too hard for most people to understand and use, let alone to use effectively and reliably.  This is precisely the circumstances that decision management was in during the late nineties when business rules technology started going mainstream.  In 2000, we crossed that chasm by introducing natural language business rules (see Haley’s Authority).  At the same time, Blaze Advisor, now owned by Fair Isaac, was crossing that chasm using a form based approach called “Innovator”.

Similar advances will be forthcoming in analytics.  As with business rules, this will not eliminate the need for highly skilled consultants, but their criticality and marginal value will diminish as analytics becomes more effective in the hands of non-experts (and as better solutions develop in key markets, such as in credit, risk, fraud and other criminality or terrorism).

Until it’s easy, use expertise

If you are in this market and could use some help with modeling, analytics, or adaptive decision management, feel free to get in touch.  We have some excellent capabilities and partners in these areas.  We are happy to help recommend approaches or products, or simply to make referrals.  Of course, there are also highly specialized consultancies, such as Strategy Analytics that can give excellent implementation-agnostic advice. 

One thing worth noting, but only in passing for now, is Fair Isaac’s acquisition of Dash Optimization.  This  reflects the increasing trend towards broader and deeper application of technology within credit decisioning.  it is also a response to the decline of scoring and the increasing need for decision optimization, which is a broader subject than decision management, with or without predictive analytics and adaptation. 

Nonetheless, optimizing portfolios will not optimize profits if the scores used are not reliably correlated with probabilities.

It is also interesting how Ilog and Fair Isaac continue to converge from a technological perspective. 

In one of his examples on “The End of Intuition”, a computer program using six variables did a better job of predicting Supreme Court decisions than a team of experts.  He focused on the fact that the program “discovered” that one justice would most likely vote against an appeal if it was labeled a liberal decision.    By discovered we mean that a decision tree for this justice’s vote had a top level decision as to whether the decision was liberal, in which case the program had no further concern for any other information. 

Credit lemmings

Mr.  Ayres went on to rationalize such simplistic criteria as indicating that humans tend to underestimate the importance of key variables, in this case the disdain of the Supreme Court for the Ninth Circuit.  Note that Mr. Ayres rationalization is closer to the truth than the model, but the model was good enough to do better than experts at predicting incomes.  Judge for yourself whether you are comfortable with his conclusion that we should therefore trust such models, even if they lack depth or intuitive aspects, to make decisions for us. 

Fair Isaac and others in the predictive analytics / decision management space would have us commit our fortunes to such models, of course.   That is their position and it is not completely without merit.   But the lemming approach of believing in models is behind the credit crunch and mortgage implosions.   Mr. Ayres mentioned this but not what had gone wrong with his thesis.  Maybe in the book…

When models fail

 Things change.  And models are never perfect in the first place.

1. One model does not best fit all.
2. Reality is not stationary.
3. There are special cases.

How many models do you need?

As an example of one model not fitting all, consider how many different kinds of people are there.  There might be one kind of person, in which case a score, perhaps even a FICO score, would be enough to describe how good or bad that person is, such as concerning credit worthiness.  If there are several kinds of people, we need to identify what kind of person is at hand and apply a model that is specific to such people.  If we don’t have a separate model for different kinds of people, then some people will lay outside the assumptions that a model must make in order to have a prediction with a useful probability. 

Note that Mr. Ayres was emphatic about the need for probabilities or other statistic measures describing the uncertainty of a conclusion produced by a model.  For example, a FICO score might be better expressed as the probability of default with a range, such as the standard deviation of that estimate.  I intend to address this point in more depth separately.

So, one model does not best fit all.  If there is only one score its uncertainty should increase as the person at hand varies from the population used to develop the model, whether there is one model or many.  But how many models are needed?  Ideally, there would be one model for each different kind of person.  This brings us back to how many different kinds of people there are.  Answering this question is obviously problematic.  Failing to answer it is also problematic.  There are two things to consider:

1. Unsupervised learning techniques, such as clustering, can answer the question objectively.
2. The number of different types of people changes over time.

Models are rarely stationary

The number of different types of people changes as the demographics of a population (e.g., our customers or web site visitors) changes.  We either understand their differences and similarities or our models will underperform, especially if they do not convey any uncertainty in their recommendations.

• Predictive analytics may use clustering as an early step in modeling.
• Adaptive decision management benefits from real-time clustering.

The evolution of people and demographics over time as an organization, its markets, and the more global political economy changes over time is only one example of change over time.  For example, what caused the mortgage crisis and credit crunch?  It was less a change in people and more the dynamics of bubble growing and bursting with respect to credit risk versus return.

Predictive modeling without risk modeling is dangerous

In conversation with the director of analytics at one of the top insurers afterward, we discussed that a good modeler would have avoided much of the mortgage implosion by having segmented the mortgage market into clusters of boom and bust rather than a relatively short window that covered only the boom in real estate over the prior ten years.  In his talk, Mr. Ayres also mentioned the implosion of quant funds that take his super-crunching approach but he did not address the need for the model to conservatively predict and how this was missed by many quants.  The answer is that they did not adequate cluster market conditions, just as we need to cluster people as discussed above, into boom and bust markets, for example.  Many quants’ models did not go back far enough to consider the possibility of conditions like those during the savings and loan crisis.  

The preceding paragraphs discuss the need to know (or discover) how many different models are needed for each cluster of distinguishable circumstances.  In addition, they address how such clusters may shift in definition or relevance over time.  In principle, the need for multiple models to cover clusters is independent of the shift in models over time, even if in practice they are correlated.

Models are generally smooth or approximate

The third challenge to accurate modeling above is the existence of special cases.  In the Supreme Court, for example, some justices have well established positions on certain issues, such as abortion or federalism, for example.   Being able to override the predictive analytic model with rules is a critical improvement, not only for compliance but for incrementally improving the performance of models in ways that are operationally relevant but beyond the ability of statistics or other machine learning to discover, especially given limited data or sudden change.  

Incrementally reducing errors of models is an obvious application of rules.  As another example, consider a stock selection model that identifies a stock that is outside of a discrete boundary in which you are willing to invest.  Either you have to teach the model learning algorithm that you don’t want to consider such targets by providing lots of negative examples, or you simply tell it with a rule. 

Rules can select models or make them more precise

Rules are a very convenient and straightforward means of reducing the false positives of any model.   And introducing such rules incrementally, as false positives are discovered, is extremely straightforward.  Rules can also address situations where a decision should be positive but the model produces a negative result (i.e., false negatives).  The problem here is that it is typically more difficult to identify when a decision not to take action was wrong.  For example:

In Fair Isaac’s case, learning that the denial of a credit card or mortgage was a poor decision is unlikely. 

It’s up to us to determine whether to tighten up the model so as to minimize false positives and thereby increase false negatives or to get a good enough model and then to manage false positives downwards using rules to handle special cases.  In either case we can model false negatives upwards using rules, if we have the insight, or if the impact (e.g., lost potential) of a negative decision can be measured.

Random Trials

As I mentioned, Mr. Ayres most interesting points had to do with random trials.  The idea is that experimentation with bounded risk allows insights to be discovered, even when they are counter-intuitive.  As examples, he showed e-commerce trials that the audience agreed would fail, but where experimentation revealed excellent performance.   Although most of his examples were web-oriented, his point is irrefutable.   To quote him:

• If you are not experimenting using random trials then you are presumptively screwing up.

Champion Challenger

Fair Isaac tries to address this with its concept of champion / challenger.  This is a good, but limited approach.   Here’s how it works:

1. Come up with a new rule that makes a different decision under certain circumstances.
2. Choose a percentage of the times the rule should be applied when it is applicable
   1. i.e., # times to be executed / # times applicable
3. Choose a metric for comparing outcomes of the alternative to the original decision.
4. Measure the outcomes using the metric across a number of decisions.
5. Decide whether to use the suggested rule 0% or 100% of the time.

The problem here is that there is no learning or predictive analytic benefit.  Everything is manual.

Google Analytics versus trials

Mr. Ayres used Google’s web page optimizer and its support for trials of web ads to demonstrate the benefits of coupling random trials with adaptive decisions management (ADM).  Specifically, Google takes weights on which ad to prefer and then it adaptively decides which one is more effective.  It doesn’t stick to the original estimates, as champion-challenger does.  Google changes the weights as experience indicates.  In effect, Google facilitates experimentation by allowing low weights to be given to new ideas.  The ideas with merit will naturally evolve into preferences.

Predictive is not adaptive

ADM is focused on closing the loop between random trials and an improving decision process.  In ADM, any number of trials can be experimented with concurrently according to an initial probability distribution.   And the adaptive aspect of ADM automatically learns how to adjust that probability distribution based on outcomes.  In effect, ADM takes the next step from Enterprise Decision Management (EDM), where leaders like Fair Isaac are integrating predictive analytics with rule-based decisions. 

Unlike EDM, ADM actually closes the loop.  EDM gives us the tools, but we still have to drive the nails all on our own.

Mr. Ayres is right on the money with random trials, but recent market experience demonstrates how risky it can be to trust predictive modeling without a deep understanding of clusters and market conditions and how they change over time.

A panelist from a major UK bank later reminded us of the old financial services adage:

• Past performance is not an indicator of future results.

The bottom line:

Predictive analytics may help make good decisions but adapting makes decisions better.

• Understanding events and processes takes time
• Oracle should teach Siebel CRM about location and money
• Ontology of time in progress – amounts needed  

The core principle here is that software needs to have a common core of understanding that makes sense to most people and across almost every application.  These are the concepts of Pareto’s 80/20 Principle.  A concept like building could easily be out, but concepts like money and time (and whatever it takes to really understand money and time) are in.  Location, including distance, is in.  Luminousity could be out, but probably not if color is in.  Charge and current could be out, but not if electricity or magnetism is in.  The cutoff is less scientific than practical, but what is in has to be deeply consistent and completely rational (i.e., logically rigorous).^[2]

Angle is in

It took working with a team that is getting software to pass AP tests in physics, chemistry, and other areas (e.g., biology and economics) for me to realize that the physicists and chemists who defined the the International System of Units overlooked something fundamental not only to math and physics, but to our integration of language with visual perception. 

It turns out that understanding angles can be just as important – and is just as primitive – as understanding distance.  Consider longitude and latitude, for example.  And if you’re trying to get a computer to get college credit by passing the advanced placement test for Physics , then you need to understand angular momentum. 

Surprisingly, distance is among the SI units, but angle is not!^[1]

Getting computers to understand angles involves circles and fractions, which bring a few other concepts into our upper ontology, such as diameter, circumference, arcs, and denominators and numerators.  Curiously, angles are primitive and yet they are intrinsically fractions of circles.

A piece of a pie could be an eighth or a sixth of the whole pie.  The angle of either such piece would be 30 or 45 degrees, since there are 360 degrees in a circle (assuming the pie is round).

What is an angle?

Here’s what Wikipedia has to say on the matter:

In geometry and trigonometry, an angle (in full, plane angle) is the figure formed by two rays sharing a common endpoint, called the vertex of the angle. The magnitude of the angle is the “amount of rotation” that separates the two rays, and can be measured by considering the length of circular arc swept out when one ray is rotated about the vertex to coincide with the other (see “Measuring angles”, below). Where there is no possibility of confusion, the term “angle” is used interchangeably for both the geometric configuration itself and for its angular magnitude (which is simply a numerical quantity).

And another definition says:

A shape formed by two rays sharing a common endpoint or two lines that intersect. An angle has one vertex and two sides.

So before we are done with all this our ontology will have to include fundamental geometry, including rays, endpoints, vertices, length (which we’ve previously covered), lines, and intersection.  Here we are going to deal with nothing more than how angles are measured and pi.

Angles are measured in degrees or radians.  Radians may seem strange at first but they are harmless.  The ratio of a circle’s circumference to its diameter is the same, regardless of its radius (which is half its diameter, of course).  This is really no big surprise, since the ratio of a square’s circumference to the length of its sides is constant, too.  Unlike a square, where the ratio is the number of sides (i.e., 4), the ratio of the circumference of a circle to its diameter is an irrational number, called pi. 

Is pi a fraction?

Most of us know that pi is approximately 3.1415.  It is also true that the digits after the decimal point continue forever without a repeating pattern.  Numbers that cannot be expressed with a mantissa that is either finite or repeating are called irrational numbers.  If a real number is not irrational then it is rational.  There are also complex numbers that mix real and imaginary numbers (involving the square root of negative one), but we’ll skip those for now, OK?

Rational numbers are not irrational

Rational numbers are normal fractions, one integer over another.  Any rational number that is not equal to an integer has either a limited mantissa or a repeating pattern.  For example, the mantissa of ½ is a single digit, as in 0.5.  The mantissa of ¼ is two digits, as in 0.25.  The mantissa of 1/3 is an infinite number of digits, all of which are 3, as in 0.33333….  The mantissa of 1/7 is a little more interesting.  It is a repeating pattern of 6 digits, as in 0. 142857142857142857142857142857142857…

So our ontology of numbers includes both rational numbers (i.e., fractions) and irrational numbers and pi is an irrational number.  But recall that rational numbers and irrational numbers are disjoint.  That is, a fraction cannot be an irrational number.  So pi cannot be a fraction!

Ratios are not fractions

Fractions are the ratio of one integer to another.  Not all ratios are between integers, however.  In fact, some ratios are between amounts, such as an amount of distance per an amount of time (e.g., miles or kilometers per hour).

Pi is not a fraction because the circumference of a circle is not an integer.  But more generally, a fraction is a ratio but not all ratios are fractions.

Angular units

An angle of one radian covers an arc of circle that is equal to the circle’s radius in length.  Since the diameter is twice the radius in length and the circumference of a circle is pi times its diameter, there are 2 times pi radians in a circle.

A degree is also an angle.  More precisely, there are 360 degrees in a circle.  And each degree equals 60 seconds of an arc.  And each second of arc has 60 minutes.  Thus, a second of arc is 1/3600^th of a circle.

Acute and other angles

So, a radian is a little more than 57° and both a radian a degree are acute angles.  All of these units of angle are therefore acute, which means that they are less than a right angle (which is 90°).  An angle of 180° is called a straight angle and an angle between a right angle and a straight angle is called an obtuse angle.  And an angle that is more than a straight angle is called a reflex angle (they are commonly used to express a heading or bearing).

Pi at last

So, pi is a positive, irrational number and the ratio of any circle’s circumference to its diameter, as shown below.

Note that e is Napier’s constant – the base of the natural logarithm (i.e., ln(x) = log_e(x)).  It is another positive, irrational number that is distinct from pi.  These are, practically speaking, the only two irrational numbers most people will ever come across in math or physics.  In fact, most people know pi but never use it.  Relatively few know or use e, but you need to know the natural logarithm in mathematics class or if you hope to get college credit for physics with high marks on the AP test.

Of course, they get to demonstrate this intelligence only because they are not shot up or blown to bits within seconds!

Here’s where AI in games is headed (from the perspective of a behavior leader in serious games):

1. Real AI is being embedded with gaming milddleware to animate non-player characters (NPCs) with much higher human-like fidelity.
2. Human performance modeling, including emotion, is being enhanced within the cognitive models of CGFs and will show up next in NPCs.
3. Crowd animation is becoming much more sophisticated and moving to center stage in role-playing games (RPGs).

The first step involves embedding a rule-based or similar cognitive architecture in game AI middleware.  These days, game AI middleware is pretty much about path planning and finite-state machine (FSM) modeling of agent behavior.  That is, it is not knowledge based and has all the limitations on context-sensitive and complexity that you would expect if you were hard-coding in C++ (Java is out since the middleware has to support the Xbox and the Playstation, as well as Microsoft Windows and Linux, in some cases.)  The FSM approach has been played out to its limit with hierarchical task networks.

For a look at some of this software, see:

• AI-Implant from Presagis
• Gamebryo from Emergent Game Technologies
• kynapse from kynogon (acquired by Autodesk)
• Jupiter from Touchdown Entertainment
• DI-Guy from Boston Dynamics

From our standpoint, Presagis is particularly well positioned to transition from serious games to entertainment.  We also think their simulation, 3D terrain, and flight dynamics capabilities are by far the best for our autonomous pilots in our military markets.  (We might have to plug-in to Microsoft Flight Simulator in the consumer market, though!)

Autonomy in gaming and robotics

Parenthetically, our agents can also fly unmanned aerial vehicles (UAVs), but for those most part those are teleoperated rather than autonomous.  Boston Dynamics BigDog is one example of an autonomous unmanned ground vehicle (UGV) that will have our kind of cognition on board in the surprisingly near future.  Boston Dynamics apparently sees the same synergy in autonomy across gaming and robotics that we do.  They are also teamed up with Presagis using DI-Guy with AI-implant. 

If you are interested in this kind of synnergy, you might also want to take a look at the following robotics middleware:

• URBI from GostAI
• Microsoft Robotics Studio

If you take a close look at the “behavior modeling” of these and game AI software you will see many conceptual similarities.

Short term AI for NPCs

Bringing rule-based cognitive capabilities to NPCs involves taking our AI technology and replacing or augmenting the FSM/HTN approaches of existing middleware is the critical step forward for AI in games.  This means C++ in order to support the Xbox and the Playstation, if not Microsoft Windows and Linux, too.  Our technology tends to be a little too heavy for the massive crowds that kynogon addresses (check out these videos).  Our technology is more capable of moderate crowds, but it is really targeted at teams of agents at high fidelity.  And our technology is perfect for high-fidelity human animation.  Check out this (and other) videos from DI-Guy, for example.

Aside from the post I referenced above, these audio/video clips can give you some insight on how much game developers are commited to AI and how hard they are working to push the limits of their current approaches:

• A podcast by a developer of the AI in Halo III.  (I too like the NPCs fighting to drive the HumVee.)
• A video with the developer of the AI in Assassins’ Creed. 

The Assassin’s Creed scenario is close to ideal for today’s cognitive AI.  You have a small crowd of agents of long-lived NPCs where the quality of the game emerges from the reality and complexity of interactions with them.  It has a long way to go but this game clearly broke ground.

The basic approach is to take the lessons from military CGFs, especially concerning cognitive and human performance modeling, including emotion, and raise the level of everything from path planning to behavior up to the knowledge-level.  If you take a close look at some of the game AI middleware, you will see that NPCs actually throw out rays to “see”.  In a few years (a game cycle or two), NPCs will just drop into games and perceive their environment with the same vision that camera and laser equiped robots use.  And they will figure how to get from point A to point B the same way you and I do.

Going into all the details of how this is unfolding is too much for me at the moment, but here are few more references on where the military folks are coming from:

• Stottler Henke tried (apparently unsuccessfully) with the FSM/HTN approach in SimBionic.
• Chi Systems tried (apparently unsuccessfully) with a proprietary cognitive architecture in iGEN.
• Micro Analysis & Design remains strictly military  CGF emphasizing stress/fatique/hunger/training more than cognition or emotion.
• Soar Technology also remains strictly military CGF with the strongest emphasis on cognition and some emotional modeling.

Note that each of these companies does more than “intelligent” agent work, I am only commenting on those aspects here.

Realistic Avatars

As more human behavior (including performance and emotion) is modeled, facial animation, including speech will become key.  I enjoy CrazyTalk from Reallusion but image metrics seem to be the facial animation leader for current games.   Also, check out Volker Blanz.

If you find this interesting, nothing that I have seen compares to the work of Stephen Stahlberg.  It will take a lot of horsepower before the agents that we interact with will be as personified as the video behind his face shown above, but will happen.

Intelligent Agents

As compelling as synthetic humans may become, their ability to engage us and be of use to us will require more the complex behavior, it will require learning and problem solving that current rule technology does not address.  Some of the cognitive architectures such as Soar and ACT-R provide these capabilities, the former on a rule-based platform.  Soar is an interesting technology with a good rule engine.  It may not be the exact approach to embed in AI middleware, but it is certainly a guidepost on the way towards intelligent autonomous agents, whether they be on screen or robots.

Recently, Nova recorded a few minute video in which he makes three short-term predictions:

1. Yahoo’s indexing of RDF will start the Semantic Web 3.0 arms race involving Google and Microsoft.
2. The web will transition from pages to linked data. 
3. Facebook “has to compete” with Google.

Nova was a little on the spot in the video.  Personally, I liked his “the web becomes a database” comment more than the Berners-Lee reiteration of linked data.  The notion of the entire web being a database is the right perspective on the semantic web (i.e., RDF), in my view.  Linked data is boring (try the Tabulator if linked data excites you.)  The action (and opportunity) is doing something with it!  When asked about ten years out, Nova displayed more of his deep insight and vision, however.  (See below.)  The truth is, beyond his first one, Nova was a little on the spot.  (See for yourself in the video.)

I love the pithy #3 that he decided to throw in there.  He did not invent that on the spot but found his legs just before being asked about longer term vision.   It makes sense, of course.  Google’s attacking with Open Social (so is the rest of the world including all the bookmarkers and even Nova’s Twine).  Facebook has to shift direction and the only target big enough given its size is search and advertising.

In his longer term vision he mentions the intelligent web that reasons and helps make decisions.  

This is where the battleground is for artificial intelligence and Semantic Web 4.0 (his term for the 4th decade of the web starting circa 2020).

Personally, I think natural language should have been in his first three.  Powerset will demonstrate that and all the action around Reuter/Clearforest/Calais (which he mentions and expects Google to compete with) indicate that natural language is critical to populating the semantic web (of course we have the database approach of DBpedia and Freebase, too).  In general, people are not going tag sentences or paragraphs.  Machines will.  The only RDF people are going to add are meta-tags at the page level for search engine optimization given Yahoo’s move (and the expected response from Google that Nova mentions.)

Certainly, natural language understanding is a prerequisite for the Semantic Web 4.0.  We will be talking more and typing less long before then.

Learning from the Future with Nova Spivack from Maarten on Vimeo.

In this article I hope you learn the future of predictive analytics in decision management and how tighter integration between rules and learning are being developed that will  adaptively improve diagnostic capabilities, especially in maximizing profitability and detecting adversarial conduct, such as fraud, money laundering and terrorism.

Business Intelligence

Visualizing business performance is obviously important, but improving business performance is even more important.  A good view of operations, such as this nice dashboard^[1], helps management see the forest (and, with good drill-down, some interesting trees). 

With good visualization, management can gain insights into how to improve business processes, but if the view does include a focus on outcomes, improvement in operational decision making will be relatively slow in coming.

Whether or not you use business intelligence software to produce your reports or present dashboards, however, you can improve your operational decision management by applying statistics and other predictive analytic techniques to discover hidden correlations between what you know before a decision and what you learn afterwards to improve your decision making over time. 

This has become known as decision management, thanks to Fair Isaac Corporation, but not until after they acquired Hecht Nielsen Corporation.

Enterprise Decision Management

HNC pioneered the use of predictive analytics to optimize decision making.  Dr. Nielsen formed the company in 1986 to apply neural network technology to to predict fraud.  The resulting application (perhaps it is more of a tool) is called Falcon.  It works.

In 2002, Fair Isaac acquired HNC (for roughly $800,000,000 in stock) to pursue a “common strategic vision for the growth of the analytics and decision management technology market”.  But shortly before the merger, HNC had acquired Blaze Software from Brokat for a song following the Dot Bomb of October, 2000 – a month before 9/11.  This gave HNC not only great learning technology but, with a business rules management system (BRMS), the opportunity to play in broader business process management (BPM), including underwriting and rating (which is highly regulated), for example. 

Of course, the business rules market has since become fairly mainstream and closely related to governance, risk and compliance (GRC), all of which were beyond the point decision making capabilities of either HNC or Fair Isaac before both these transactions.

Once Fair Isaac had predictive and rule technology under one roof, bright employees such as James Taylor, coined “Enterprise Decision Management”, or EDM for short.

Predictive Analytic Sweet Spots

Before it merged with Fair Isaac, HNC’s machine learning technology was successful (meaning it was saving tons of money, not just an application or two) in each of the following business to consumer (B2C) application areas:

• Credit card fraud
• Workmen’s compensation fraud
• Property and casualty fraud
• Medical insurance fraud

Clearly fraud, across insurance and financial services is a sweet spot for decision management.  Today, that includes money laundering and, in general, any form of deceit, including adversarial forms, such as involving terrorism.

HNC also moved into retail and other B2C areas, including:

• Targeting direct marketing campaigns
• Customer relationship management (CRM)
• Inventory management

Some of the specific areas in marketing and CRM included:

• “Up-selling” (i.e., predicting who might buy something better – and more expensive)
• “Cross-selling” (i.e., predicting which customers might buy something else)
• Loyalty (e.g., customer retention and increasing share of wallet)
• Profitable customer acquisition (e.g., reducing “churn”)

The inventory applications included:

• Merchandizing and price optimization
• SKU-level forecasting, allocation and replenishment

Predictive Analytic Challenges

The principle problem with predictive analysis is the care and feeding of the neural network or the business intelligence software.  This involves formulating models, running them against example input data given outcomes, and examining the results.  For the most part, this is the province of statisticians or artificial intelligence folk.

A secondary challenge involves the gap between the output of a predictive model and the actual decision.  A predictive model generally outputs a continuous score rather than a discrete decision.  To make a decision, a threshold is generally applied to this score. 

1. Yes or no questions are answered by applying a threshold to a score produced by a formula or neural network to determine “true” or “false”.
2. Multiple choice questions are answered using a predictive model per choice and choosing the one with the highest score.
3. More complex decisions are answered as above using a predictive model that combines the scores produced by other predictive models.

In general, especially where decisions are governed by policy or regulation, predictive models and decision tables are combined with rules using one of the following approaches:

1. More complex decisions are answered as above using predictive models that are selected by rules in compliance with governing policy or regulations.
2. More complex decisions are answered using rules that consider the scores produced by predictive models in compliance with governing policy or regulations.

In general, governance, risk and compliance (GRC) requires rules in addition to any predictive models.  Rules are also commonly used within or to select predictive models.  And special cases and exceptions are common applications of rules in combination with predictive models.

Scorecards

A simple case of defining (or combining) predictive models is a scorecard.  The following example shows a scorecard from Fair Isaac’s nice brochure on predictive analytics that could be part of a credit worthiness score:

Fair Isaac is the leader in credit scoring, of course.  Their FICO score is the output of a proprietary predictive model.

The following example shows how Fair Isaac’s predictive model is combined with other factors in the mortgage industry (click it for a closer look):

Note all the exceptions and special cases spread throughout this scorecard!

This explains why business rules have been so popular in the mortgage industry.  Pre-qualifying and quoting across many lenders clearly requires a business rules approach (which explains why Gallagher Financial embedded my stuff in their software a decade ago).  Even a single lender has to deal with its own special cases and the bigger the lender the more there are (which is why Countrywide Financial>^[2] developed its own rules technology, called Merlin, decades ago).

Decision tables

For anything but the simplest decisions, the results of predictive models are considered along with other data using rules to make decisions.  In some cases, these rules are simple enough to fit into a decision table (or a decision tree rendered as a table) such as the following:

Tables like the one on the left can be used during underwriting to determine what variables are appropriate for gauging the risk of death covered by a life insurance policy.  This demonstrates that rules (in this case, very simple rules) can be used to determine which predictive model (or inputs) to consider in a decision. 

Tables like the one on the right correspond to decision trees and can be used instead of scorecards to set the base premium for auto insurance.  Additional rules typically adjust for other factors like driver’s education classes, driving record, student drivers, and other special cases and exceptions.  This is similar to the use of notes in the mortgage pricing sheet shown above.

The point is that real decisions are not as simple as a single predictive model, a scorecard, or a decision table.  And once these decisions are defined and automated using any combination of these techniques, improving those decisions can seem overwhelming complex (just from a technical perspective!)

Predictive analytics is not enough for EDM

Enterprise Decision Management (EDM), discussed above, is all about this multi-dimensional decision technology environment (scorecards, decision tables, and rules) but also about bringing statistical and neural network technology in to improve the decision making process more easily and less manually or subjectively.  The Fair Isaac brochure referenced above, for example, has some nice graphics showing statistical techniques (such as clustering) and graphs showing interconnected “nerves”.

There are several aspects of decision making that not even magically successful machine learning will eliminate, however:

1. The requirement to comply with governing policies or regulations.
2. Special cases that cannot be learned for various reasons, including:
   • Limitations on the number of variables used in predictive analytics.
   • Poorly understood, non-linear relationships in the data
   • A lack of adequate sampling for special cases
   • A need for certainty rather than probability
3. Exceptions that cannot be learned, as with special cases.

Of course, special cases and exceptions are common in both policy and regulation.  For examples, consider policies that arise from contracts or customer relationships or the evolutionary nature of legislation, as reflected in the article on the earned income tax credit.

Rules are not enough for EDM

On the other extreme, commercial rule technology has not been capable of adaptively improving decision management.  In fact, except when they are modified by people, the use of rules in decision management is completely static, as well as entirely black and white.  There is no learning with any of the business rules management systems from leaders like Fair Isaac, Ilog, Haley, Corticon, or Pegasystems.

Amazingly, there are no mainstream rule systems today that deal with probability or other kinds of uncertainty.  Without such support, every rule in the tools from the vendors previously mentioned is black and white.  This makes them very awkward for applications such as diagnosis.  And all decision management applications, including profit maximization and all forms of fraud detection, are intrinsically diagnostic.  Prediction results in probabilities!

The earliest diagnostic expert systems were developed at Stanford.  One used subjective probabilities to diagnose bacterial infections.  Another used more rigorous probabilities to find ore deposits (it more than paid for itself when it found a $100,000,000 molybdenum deposit circa 1990!).  These applications were called MYCIN and PROSPECTOR, respectively.

This seems shocking really, since the technology of these systems is well-understood and technically almost trivial.  The truth is that the Carnegie Mellon approach to business rules has won because it dealt with “the closed-world assumption”, which means that it could handle missing data better.  But CMU’s approach was strictly black and white.  Stanford was left in the dust commercially after the success of OPS5 at Digital Equipment Corporation and the commercialization of expert systems at Carnegie Group, Inference Corporation, IntelliCorp and Teknowledge left uncertainty in the dust during the mid-eighties.  Neuron Data, which became Blaze, followed the same trail away from the uncertain toward the black and white of tightly governed and regulated decisions.

With nothing but black and white rules, EDM leaves it up to people to adapt the decisions.  Sure they can use predictive analytics, but there is no closed loop from predictive analytics involving rules.  Any new rules or any changes to rules follow the stand-alone business rules approach.

Innovate for Rewards with bounded Risk

One problem with black and white rules technology is that it forces you to be right.  This stifles innovation.  Ideally, you could formulate an idea and experiment with it at bounded risk.  For example, you could say “what if we offered free checking to anyone who opens a new credit card account with us” and test it out.  You don’t want to absorb the cost of thousands accepting your offer only to lose more on checking fees that you gain through credit card fees.  So you indicate how often or how many such offers can be made. 

Not surprisingly, this approach is tried-and-true.  It’s most common form is the champion/challenger approach.  Fair Isaac has been “championing” this approach for some time (see this from James Taylor).

But how do you close the loop?  How do the rules learn when this new option should be used to maximize profit?  The fact is, they don’t.  People do it using predictive analytic techniques and manually refining the rules.

The problem, once again, is that the rules do not learn and that their outcomes are black and white.  The rules do not offer a probability that this will be a profitable transaction.  And they do not learn whether a transaction will be profitable over time, either.  That’s up to the users of predictive technology and managers.

Adaptive Decision Management

 Adaptive Decision Management (ADM) is the next step in EDM.  In ADM the loop between predictive analytics and rules is closed.  At a minimum this involves learning the probabilities or reliability of rules and their conclusions.  This learning occurs using statistical or neural network techniques that can be trained, optimized, and tested off-line and – if your circumstances allow – even allowed to continue learning and adapting and optimizing while on-line.  For example, advertising, promotional (e.g., pricing) and social applications almost always adapt continuously.  Unfortunately, none of them use rules to do this yet, since the major players don’t support it!

Innovation and ADM

The adaptation of rule-based logic brings new flexibility and opportunity to the use of rules in decision management.  Adding a black and white business rule requires complete certainty that the rule will result in only appropriate decisions.  Of course, such certainty is a high hurdle.  Adaptive rules have a relatively low hurdle.

With adaptive rules an innovative idea can be introduced with a low, or even a zero probability.  As experience accumulates, the learning mechanism (again, statistical or neural) determines how reliable the rule is (i.e., how well it would have performed given outcomes).  The technology can even learn how to weight and combine the conditions of rules so as to maximize their predictive accuracy.  Without learning “inside a rule”, the probability of the rule as a whole may remain too low to be useful.  And, unlike a black box neural network, the functions that combine conditions and the probabilities of rules are readily accessible, whether for insight or oversight.

The overall impact of adaptive rules is that you can put an idea into action within a generalized, probabilistic champion/challenger framework.  And using techniques such as the subjective Bayesian method used in MYCIN or other more rigorous techniques as in PROSPECTOR, more patterns can be considered and leveraged with the continuously improving performance that EDM is all about.

The advantages of ADM include:

1. the improving performance of EDM
2. faster and more continuous improvement versus manual EDM
3. a generalized approach to champion / challenger using probabilities
4. better predictive performance than manually maintained scorecards or tables
5. improved performance over black and white EDM by leveraging innovation adaptively

Although they haven’t told me about it explicitly, I would expect Fair Isaac to move in this direction first among the current leaders given their EDM focus.  I would not be surprised to see business intelligence (BI) vendors, perhaps SAS,  move in this direction, too.  I know it will happen since we are already working with one commercial source of adaptive rules technology.  Unfortunately, Automata is under NDA about their approach for now, but stay tuned…  In the meantime, if you’re interested in learning more, please drop us a note at info at haleyAI.com.  And if you see any issues or good applications, we would love to hear them.