Harvard Business Review has just published a blog — A Process for Human-Algorithm Decision Making by Michael C. Mankins and Lori Sherer (blogs.hbr.org/2014/09/a-process-for-human-algorithm-decision-making) that could be a source of confusion. In any case, it provides a nice teaching opportunity. Many in the process area are, at the moment, interested in how to blend conventional process work with analyses of decisions. And some of us have been advocating the use of business rules as the way to capture and structure advanced decision support. To clarify this, its worth considering how rules relate to algorithms.
Decision Management, Business Rules, IBM's Watson and Cognitive Decision Support all derive from work done in AI labs in the Eighties and tested in expert system applications developed in that same timeframe. It might be worthwhile to get a little technical about what advanced knowledge capture and decision support is really all about. It's about solving problems that can't be solved using conventional algorithms.
An algorithm describes a step-by-step for accomplishing a task. To create an algorithm, one begins by defining a search space, which describes all possible steps one might take in an effort to accomplish the task. A search space is usually described by means of a branching tree structure. One begins somewhere and it faced with possibilities. Then one considers each possibility and that option, in turn, produces other possibilities. If you think of a game like Tick-Tack-Toe, you can begin by placing an X (or a O) in any of nine locations. Once you have chosen one of the nine locations and made a mark, you opponent has 8 locations left and can place a mark in any one of those. The goal is to get three similar marks in a horizontal, vertical or diagonal line. The search space for this problem isn't too large, but once you work it out, with all the branches, then you are able to examine the resulting tree and see if their is a path that the first player can follow that will lead to the desired result. As it turns out, there isn't. Someone can win the game if their opponent makes a mistake, but if both players are on their toes, neither party can win. Suffice to say that if you played against a computer, the computer would always either get a Draw, or it would Win if you made a mistake. We can be certain of this because we have examined the complete space and we can see that their is no path (algorithm) that guarantees a win.
Now what about a game like chess? Chess is played on an 8×8 board with 16 pieces on a side, and different pieces are allowed different types of moves. A bit of calculation shows that the search space for a chess game would be huge. In fact, it crosses a line and is so large than it could not possibly be searched. (There would be more branches than their are atoms in the universe. Such a search space could not be stored in any imaginable computer memory AI practitioners refer to such large spaces as NP-hard — Non-deterministic Polynomial-time hard.) Thus, we can never be sure whether or not there is an algorithm that would guarantee that white would always win (or could never be assured of a win) in chess.
The major AI development of recent decades involves ways to deal with NPHard problems. The answer is Heuristics in combination with search algoritums that settle for approximate solutions. A heuristic, for our purposes here, is a rule. It states that: If X is true, then something else is true or false. And it the most sophisticated systems it assigns a probability to the rule. An algorithm — which AI practitioners tend to refer to as a search engine — sorts through rules based on user inputs and reaches any conclusions that can be logically arrived at. This approach can not guarantee a correct outcome — these results are not like those produced by algorithms in classical computing — but it can often suggest an outcome with a high probability of being correct.
Now consider a practical problem: A bank board wants to determine if it should make a hundred million dollar loan to a sovereign nation. This problem is at least as hard as chess — there are lots of variable that are constantly changing. There is no possible algorithm that will define a step-by-step way of calculating the right answer. We can, however, interview lots of experts and capture the rules of thumb (heuristics) they use as they think about the problem. Those heuristics, with assigned confidence factors can generate possible solutions with various probabilities. This can certainly help guide a human expert when he or she is faced with such a decision.
Humans wouldn't be able to sort through the thousands of rules and their associated probabilities to reach a solution in a reasonable time, but computers can. In effect, software applications incorporating AI techniques (search algorithms) and data bases of business rules can support human decision makers.
Obviously the thinking that has gone into figuring out how to deal with large search space problems is very similar to the thinking that has recently gone into Case Management problems. How do we analyze business process flows when the process keeps changing and is very complex. The answer is that we can't model the problem as accurately as we could if the problem were simple and stable, but we can come up with a good approximation. And a part of that approximation will probably incorporate decision support elements that will rely on business rules to indicate likely decisions.
Returning to the Blog that kicked off this comment: There are many processes in which humans can work with algorithms to make decisions. More important, however, we are now beginning to deal with problems that can't be solved with traditional algorithms, but that can benefit from business rules and computer-implemented search engines. Increasingly process analysts will use cognitive techniques to capture rules and create applications that can support very complex human decision-based activities.
[For a detailed explanation of the basic techniques described here, check Harmon and King, Expert Systems: AI in Business. It's a book from the Eighties, but the explanations are very good and this is basic stuff that's not out of date yet.]