In-Depth

I, Smartapp

• By Stephen Swoyer
• December 1, 2005

The Big Idea

Adaptable Applications

• Even the most elaborate problem-solving systems can be reduced to a rules repository (or knowledge base) and an inferencing engine.
• Many of the problem-solving techniques that power adaptive learning and rule-based inferencing systems have gone mainstream.
• There are a couple of vexing impediments to the development of smart apps in most enterprises—namely, a lack of expertise and time.

Some of the techniques first pioneered by AI researchers (and commercialized by the expert systems pioneers of the 1970s and 1980s) have been used to produce highly adaptive applications or systems. Many expert systems— which can range from the simple (true/false logic) to the complex (true/false logic in tandem with fuzzy logic)—might be called smart, because they're able to make reasonable inferences about the behaviors and preferences of their interactive users and adjust their application presentation or strategic guidance accordingly.

More to the point, the problem-solving capabilities that power the expert systems of today are explicitly portable, because they aren't based on domain-specific expertise. With a few exceptions, even the most elaborate of expert or problem-solving systems can be reduced to two components: a rules repository (or knowledge base) and an inferencing engine.

AI and expert systems concepts and methods aren't just the province of academics or highly specialized commercial ISVs anymore. Many of the problem-solving techniques that power artificial intelligences or expert systems (adaptive learning and rulebased inferencing capabilities) have gone mainstream. Most codejockeys are at least familiar with pathfinding, decision trees and rule-based inferencing, if only as vague holdovers from otherwise long-forgotten CompSci AI seminars.

But some, like Steve Berczuk, a programming consultant and co-author, with Brad Appleton, of Software Configuration Management Patterns, are using these techniques in their ongoing work. "I once worked on an application that used genetic algorithms for scheduling—or rather for optimizing schedules. I'd classify it as intelligent in the sense that it came up with answers comparable—or better—than a person might have," Berczuk explains. "On the other hand, the way that genetic programming works is by trying different things and picking the best options in an efficient way. So it was intelligent, but achieved it by doing lots of small simple things, and correcting course."

Others, such as Stefan Schmiedl, a programmer with risk management specialist Approximity, are starting to experiment with smart applications in production environments for the first time. "We actually have [used] adaptive techniques to improve prediction in a prototype currently under heavy development," he confirms.

Not a trend, yet
Dan Massey, chief technical architect with Borland Software, says smart application development isn't a widespread trend—yet. "I haven't seen that much of it coming in from customers in terms of requirements [for app dev tooling]," he confirms.

"They're doing a lot of imperative coding, and some of them are kind of finding new ways to solve traditional problems, but I've actually had to force rules [inferencing] on people," he says.

There are impediments to the development of smart apps in most enterprises—a lack of expertise and a lack of time. Codejockeys aren't AI researchers, nor do most rank-and-file programmers have much experience with AI or expert systems concepts and methods.

And even if they did, Massey points out, most projects don't give codejockeys enough time to build the infrastructure services—such as a robust inferencing engine—that are needed to enable smart apps in the first place. "The amount of time to do it is more than most developers are given to devote to a project," he points out. "I think a lot of developers could embrace this stuff and go forward with it, but implementing it in the real world, they're not given the time to go do that." That's changing, however.

One emerging model for building potential smart apps is the business rules approach, which to some extent is an offshoot of expert systems research. At a basic level, BRA describes a model in which (often declarative) rules determine outcomes; in their more sophisticated incarnations, business rules management systems (BRMS) use inferencing engines to decide which rules are called for and why. In this respect, organizations can configure their BRMSes to automatically trigger decisions that might otherwise have to be made by a human being—such as a business analyst. A host of vendors market BRMSes, and at least one shrinkwrapped player—Microsoft—has embedded a rules engine in a commodity product (Biztalk Server).

The upshot, says BRA proponent Ron Ross, co-principal of consultancy Business Rules Solutions and executive editor of BRCommunity.com, is there's a market for off-the-shelf BRMSes. And while there is still a host of issues associated with these products—the codification and management of rules, the rapid rate of change (and accompanying evolution of rules) in many organizations, the human resistance to divulge privileged knowledge—they can eliminate a lot of the complex and timeconsuming coding that would otherwise preclude the construction of smart apps in many organizations.

"There have been, probably out of necessity, a lot of do-it-yourselfers in the past, but...the momentum to acquire off-the-shelf capabilities has grown tremendously, and very few people are undertaking to develop their own [BRMSes] these days," said Ross in an interview last year. "At the very least, there are a couple of open-source places that people go so they won't code their inference engine or rules engine from scratch."

The BRAdescribes a very specific paradigm— in this case, using declarative rules designed to appeal to business users—but the concepts on which it's based (rules and inferencing algorithms) have much broader applicability. As a result, even skeptics say it's possible to use a combination of rules (declarative or otherwise) and inferencing algorithms to construct highly adaptive, quasi-intelligent apps.

And given the availability of commercial, off-the-shelf and open-source rules management systems, it's easier than ever. Although many question the wisdom of doing so. "It's definitely possible," concedes Ron Jeffries, a senior consultant with The Cutter Consortium's agile software development and project management advisory service. "I think that for most purposes, an inferencing engine is probably overkill. It's yet another instance of a framework looking for something to do. I'm sure there are exceptions. But I don't run into them."

People unnerved
In a certain sense, it comes down to a fundamental human need, argues Borland's Massey. "People...want some [application intelligence], but they also want it to be deterministic," Massey says. "Consider Bayesian stuff. In theory, you could test for it, you could trend, expecting it to pass or fail. But even though you know it is deterministic somewhere down in the guts of it, it could still seem like magic. That unnerves many people. They want to be the ones who are in control."

Programmer Schmiedl says he has firsthand experience with this phenomenon. His company incorporated adaptive learning capabilities into one of its risk-management software prototypes, designed to increase the predictive capabilities of the prototype, Schmiedl says. At the same time, improved predictive insight does entail a cost of sorts—at least from the perspective of some potential users.

"I demo'ed some of the techniques to a client about a year ago, who replied that he really felt uncomfortable with a system that unpredictable. He valued reproducibility over evolution," Schmiedl explains.

Schmiedl's experience is seconded by Mary Crissey, a marketing director with data mining and statistical analysis powerhouse SAS Institute. To some extent, Crissey concedes, it might be possible to use rules and inferencing algorithms to automate the actionable insights that are unearthed via data mining or information analysis. After all, she says, data mining and information analysis is undergoing a renaissance of sorts, thanks to the ability to process unprecedented amounts of data. But empowering the application to make decisions of this kind isn't a direction in which SAS is going, she says.

"It's more of an alert [model] right now. It's showing you things that you might not have seen because of all of the possible combinations. That's where we are," she explains. But couldn't SAS automate information analysis via rules, inferencing or genetic algorithms, if only on an industry-specific basis?

Certainly, she says, SAS can hardcode trending information to custom-tailor data mining or text mining solutions for specific customer environments. As for automation via quasi-intelligence, Crissey is more reserved. "I would really want a human being [involved]. I think most of our customers expect that [the output of] analysis will be interpreted by a human being. The point is that the machines are not running by themselves. I am strongly in favor of the human participant being a valuable part of the decision-making process."

Unsure smart apps will be smart enough
In many cases, of course, artificial decision- making can be a good thing. E-commerce applications use rules and inferencing to trigger discounts, suggest additional or alternative items, and provide incentives (free shipping, anyone?) based on a customer's online buying patterns. Healthcare, insurance and financial services firms—or any company that's subject to increased regulatory oversight—can use rules to more easily manage changing regulations (rules aren't hardcoded into applications but exist as declarative statements in a repository), ensure compliance with requirements or mitigate exposure to potential litigation. (A rule-based, automated loan approval process is by definition impartial.) As for HAL Corp.'s permitting its Executive Information System to trigger a merger with XYZ Corp., that won't happen.

"There doesn't even have to be any awareness for something like this. [The computer is] going through its rule base and everything it knows, and it looks at the options it's given you and which ones you've picked. Then some algorithms kick in, and it has a library of different ways to show you its data. It goes through different ways of prioritizing them; maybe it starts to learn that you always pick the second option. It learns through trial and error, and from there, you start moving toward teaching it what marketing is," he concludes. "If humans don't want machines thinking for them, maybe you might want to focus on putting in the guts so [the machines] can market their decisions to you."

ILLUSTRATION BY RYAN ETTER