[Editor's note: this is a guest post by our very own Tom Wilger, an expert in decision management and rule systems. The original version of this post can be found here, and it is republished on the BP3 blog with permission of the author]
As a decision scientist, I work in a discipline that has become known as? "Decision Management". When I look this up in Wikipedia it says that decision management "entails all aspects of designing, building and managing the automated decision-making systems that an organization uses to manage its interactions with customers, employees and suppliers." Depending on the people I hang out with, and the tools they bring to bear for problem solving, we have a tendency to limit the scope of problems that we include in this discipline. Since many of my colleagues work in the Business Rule Management System (BRMS) space, I've noticed when working with them we sometimes equate Decision Management with business rules. While it is true that many of these problems can be solved using a BRMS, there are also categories of problems that cannot be solved using business rules, or require some other techniques alongside business rules. Lately, I've been using this diagram to articulate the different categories of problems we encounter in Decision Management.
At the top of the diagram I've divided the problems based on the type of information used to make the decision. Deterministic information is based on data that we can know, or at least are assuming to know its value. Such data could be age of a person applying for an insurance policy, or the amount of product we need to produce during a production run in a factory. Stochastic information is based on data that we don't know for certain, like what the high temperature is going to be tomorrow in Albuquerque, New Mexico, or what the closing price of Exxon stock will be one week from now.
On the left side of the diagram, I've divided the problems based on types of complexity. A decision is independently complex when the outcome of one instance of a decision is not influenced by, nor influences the outcome of any other instance of that same decision. For example, when an insurance company is trying to determine if an applicant if eligible for a policy, that eligibility is not influenced by the eligibility of any other applicant. These types of problems can typically be solved using business rules in a BRMS.
We say that a decision is dynamically complex when the outcomes of instances of the same type of decision impact each other. So, if a brewery is trying to decide how to schedule an order for the production of a light beer, the start time of that production order would likely be influenced by the start time of a dark beer production order since it is usually advantageous to sequence these in order from light to dark to minimize cleaning. These types of decisions cannot typically be solved using a BRMS, but are solvable using various optimization techniques such as mathematical or constraint programming.
Decisions based on stochastic information are similar except that prior to making the decision, we must estimate the value, or the potential range of values of the stochastic variables. For instance, if an auto insurance company is pricing a policy for a 16-year-old boy with a red sports car, they would be interested in the probability that he will be involved in an accident. This estimation of a claim will likely be different than that of a 45-year-old woman driving a brown Prius. This actuarial pricing decision would still be made using a BRMS, but some of the rules would be based on first performing some type of Predictive Analytics. Although not as common, there are even dynamically complex decisions that are based on stochastic information. Hedge fund managers may use Stochastic Optimization to optimize an investment portfolio. The stochastic data in this case may be the expected future price of a stock, or some economic factor such as the expected inflation rate. These types of problems are less common since they get very large and computationally expensive to solve as the number of stochastic variables increase.
This model is by no means perfect. For instance, other types of decisions such as those made by event-based systems don't really fit into this structure. These decisions are quite different from the ones I've described in that they tend to be a sequence of simple decisions that happen over a period of time. Examples of such decisions include decisions made to detect fraud, or decisions made by autonomous vehicles. These decisions include monitoring systems that are designed to respond in real time to changes in parameters. Perhaps we can articulate these differences based on the horizon over which the decision is being made. This is something to think about.