Most of us are comfortable with making decisions. This is good because each of us makes an incalculable number of decisions every day. The most interesting and complex decisions that we make are voluntary decisions where we are able to apply discretion in why, how, and what decisions are made. However, it would be a mistake to ignore the fact that we also make an immense number of involuntary decisions every day. These range from decisions largely beyond our control like the electrical signals that are evaluated by our CNS and cause our heart to beat to reflexive behavior (remove your hand from a hot stove) and ultimately learned behavior (don’t touch a hot stove). The common pattern that unites all of these behaviors is stimulus-response (S-R) theory. Figure 1 shows a schematic of the S-R model. In this S-R model, the sense activity recognizes a change in the environment. This change is a trigger, event, or simple change of state. The act activity is an action taken in response to a particular sensation.

While S-R theory is conceptually simple, it does raise a question about what happens when a choice can be made regarding what action to take. Early thinking on the topic of event-driven architecture mimicked S-R processing by having events directly associated with actions. While this approach is extremely efficient, it is also brittle, which limits its utility in today’s IT environment where applications must be engineered for change and therefore loosely coupled. Without the ability to support a level of indirection between sense and act, there is no way to easily accommodate change. By introducing a decision node between sensing and acting, we now have clear separation of concerns and the flexibility to link any sensory event with any action, as shown in Figure 2. This enables us to refer to this modified S-R model as a decision model.

By introducing a decision node, we allow for different types of decisions. This more robust model can also emulate an S-R model simply by either always choosing the same action or defaulting the decision (such as the “else” clause in an “if-then-else” expression). However, the value of this decision model is that it recognizes that:

There is a decoupling between sensing and acting, and actions are governed by decisions.

The existence of competing alternative actions to a particular set of stimuli mean that a decision process is needed.

A decision process must take into account that available stimuli may not be sufficient or specific enough to clarify what action to take.

Decision outcomes, actions, and impact may be useful in influencing future decisions.

The ability to align specific stimuli with a particular action through a decision provides flexibility and consistency.

The act of decisioning is complex and many techniques can assist in the decision making process.

Despite the importance of decisions, we live in an action- and process-centric world. Decisions determine the potential utility to be gained, but actions are what drive kinetic utility or recognized utility. Actions (or behavior) are what define and differentiate an enterprise. Because actions can be directly tied to utility, it is easy to dismiss the importance of the decisioning. However, no action should ever be taken unless preceded by a decision. Decisioning is where context, alternatives, potential utility, objectives, constraints, and trade-offs are evaluated and a next-based action is determined. Therefore, support for comprehensive decisioning is critical because the decision is where the choice is made between competing actions. This choice can have lasting impact especially if it is strategic and this also means that decisions can have significant consequences, both positive and negative. Consequently, organizations will want to always make the best possible decisions that they can in order to maximize benefit and minimize risk over some time horizon.

Some decisions are simple and some are complex. Complex strategic decisions are often wide in scope, high in risk, few in number, and difficult to automate, and leverage inputs from many sources. Simple tactical decisions are typically the opposite; limited in scope, require few inputs, are low in risk, are large in number, and easy to automate. As decisions increase in complexity, so too does the need for analytics to support the decision making process. The point is that the decision model can be extended to include an analysis activity where the heavy lifting of evaluating alternatives is performed prior to decisioning. Figure 3 presents this as a decision analytics model.

Separating analyze from decide has distinct advantages. The primary advantage is a separation of concerns. The analyze activity is focused on understanding, quantifying, and normalizing alternatives so that a rational and informed decision can be made. It should be noted that this decision analytics model does not state any requirements regarding latency. While S-R models typically have a distinct real-time orientation, this is not the case for all decision and decision analytics models. Not all decisions that require analysis can or need to be pursued in real time. There is, however, a growing emphasis on and trend toward real-time decision analytics, so adoption of application architectures that support real-time decision analytics is appropriate although not all decisions will need to be made in real time.

When we evaluate the decision analytics model in Figure 3, it is apparent that we can improve on this model in several ways. The sense activity can be improved if we explicitly specify that a discovery activity’s whole role is to consider the relevance of new and different types of events and triggers that will have an impact on decisioning. The analyze activity also benefits from an enrichment activity that improves the understanding of context, alternatives, and additional information related to decisioning. The decide activity also benefits from an understanding of policy expressed by objectives and constraints that govern decisioning. Figure 4 improves upon the decision analytics model by adding discover, enrich, and set goals activities, which move the model toward a true reference model for decision analytics.

The discover, enrich, and set goals activities are classified in Figure 4 as “pre-decision” activities. Pre-decision activities improve the sense and analyze activities by enabling a more comprehensive analysis of events, information, and factors that will influence the decision. These pre-decision activities also improve the decide activity by defining policy-oriented objectives and constraints apriori. Objectives are goals intended to shape decisions so that an organization has targets that it aspires to achieve. Constraints are goals intended to shape decisions so that an organization operates within limits that will minimize its risk exposure legally, financially, or ethically.

These pre-decision activities are a first step in bringing a lifecycle to decision analytics. Pre-decision activities have strong bi-directional relationships with analytic decisioning because of their focus on decision improvement and the support they can provide prior to decisioning. Also, consequently, a separate set of post-decision activities complete the feedback loop. Figure 5 introduces these post-decision activities.

The post-decision activities in Figure 5 consist of evaluate, learn, and adjust activities. The intent of the evaluate activity is to assess the utility generated by an act activity and compare it with the desired utility as defined by the set goals activity. The learn activity is the capability to remember the output of the evaluate activity. The evaluate activity also factors what has been learned into its assessments so that the utility of the current action can also be compared with past actions. The role of the adjust activity is to consider the goals, decisions, actions, and what has been learned to improve performance by changing the triggers, events, analysis, and decisions. The adjust activity is where the loop is closed as in a closed loop system. The adjust activity is also one of the most complex activities that exists in this system. This is because changing policy and decisions changes actions, which will have a different impact than that to which the organization is accustomed. Changes to policy that correct errors are expected to increase utility. However, changes to policy in search of added revenue are more challenging and must be evaluated more carefully to ensure that the return outweighs the risk. Economic models are very effective at evaluating risk and return and can be incorporated in either the adjust or analyze activities. A summary of pre- and post-decision activities is as follows:

Discovery is the identification of events, objects, situations, and relationships that will have a bearing on decisioning.

Enriching is the process of incorporating content surfaced in the discovery process into the decision making process.

Setting goals is the specification of objectives to guide the decision making process.

Evaluation is the process of assessing the impact of the action taken.

Learning is the act of acquiring knowledge specific to decisions made and actions taken.

Adjusting is the act of applying knowledge gained from the learning process to improve the decision process.

It is important to note that while we have identified pre-decision and post-decision activities, we have not made any claims regarding temporal requirements for decision analytics. We do, however, expect a wide variety of use cases depending upon the analytical techniques employed that range from offline to real-time decision analytics.

Figure 5 is labeled as the decision analytics reference model. The reason for this is that this model captures the key activities and relationships that should exist within any organization that intends to address analytic decisioning both comprehensively and effectively. This decision analytics reference model primarily focuses on decisioning and how leveraging analytics to do both support and improve decisioning. The decision analytics reference model also means that consideration has to be given to application architecture. If there is an assumption that some decision analytics activities must be supported in real time, then events, messaging, state, push, and mobility must be factored into system design.

Real-time Decisioning and the Internet of Things

Real-time decisioning is an important area of investment for many enterprises. Infrastructure is now being put in place to capture data streams in real time, analyze this data, and make decisions in real time. Examples of real-time systems are everywhere. Simple real-time systems are S-R systems such as a home alarm system. More sophisticated decision analytics systems are event-based and perform some analysis before making a decision as to what action to take. An example of this would be the grocery store checkout, which generates coupons based on your purchases and frequency of visits. Even more complex decision analytics systems use feedback to adjust actions in real time. An example of this would be an automotive accident avoidance system, which monitors your distance and closing speed to an object and then applies the brakes progressively to prevent an accident. All of these real-time examples involve a subset of capabilities resident in our decision analytics reference model.

The Internet of things (IoT) is going to be very effective at connecting people and “things,” whereby a thing is an electro-mechanical device that could range from a simple sensor to an intelligent micro-processor enabled device. The utility of the IoT will be derived from its support for all person/system interactions patterns. The most interesting of these patterns will include system to person and system to system. The system to person interaction pattern will present a person with opportunities or concerns that warrant her attention. The system to system interaction pattern will need to unfold in an as-of-yet undefined way but will likely involve gateways for gathering and consolidating domain-specific information and new communication architectures, some of which will mimic high-level architecture (HLA) that was developed by the Department of Defense.

The decision analytics reference model is important because it not only identifies the significant role of analytics in decisioning, but also provides the necessary context for describing the decision analytics continuum.

The Decision Analytics Continuum

The decision analytics continuum was born out of a need to help organizations understand the various analytic techniques that they can employ to support or improve decisioning. The principles of the decision analytics reference model are to provide a generalized decision making model that also emphasizes the importance of decision improvement. This ensures continued relevance of the decision model given a changing environment and creates opportunity for vendors that deliver these capabilities and enterprises that leverage these capabilities effectively. Opportunity in this context is defined as:

Greater precision in responding to needs.

Faster understanding of changing conditions, which encourages innovation.

Improved operational efficiency due to more comprehensive understanding and rendering of organizational activities.

Better decision making.

Improved time to decision/action.

Now that we have established the importance of decisioning and the framework for decision improvement, we can explore differing analytic techniques to support decisioning. When we examine what analytic techniques support decisioning, it is useful to select criteria that will allow us to categorize these analytic capabilities. Four criteria have significant relevance in this task and include the following:

1. Decision Scope. Decision scope refers to how focused the decision is as measured by the cardinality of its alternatives or intended audience. Course-grained decisions are ones that have few choices and apply to only a few market segments (large groups). Fine-grained decisions can have many possible choices and apply to many market segments (such as markets of one).

2. Decision Execution. Decision execution refers to how much is known about the decision outcome. Deterministic decisions are ones where a particular set of stimuli always lead to the same decision. Non-deterministic decision outcomes vary based on accumulated knowledge at the time of the decision.

3. Decision Uncertainty. Uncertainty is a cornerstone of modern statistics. Analytical techniques enable us to evaluate past and present decisions as well as gain insight into how actions may influence future decisions. Since the future is not certain, understanding and quantifying the likelihood of a future event is useful to support future decision making. Collaborative decisioning, Bayesian statistics, and adaptive systems all should or do factor uncertainty into their decision making activities.

4. Decision Complexity. Decision complexity is driven by the number of factors that must be jointly considered when making a decision. The greater the number of factors (or variables) the more potential outcomes and the more complicated it is to make a decision.

Decision scope and decision complexity are closely related. Course-grained decisions tend to have less complexity and fine-grained decisions tend to have much higher complexity. Decision execution and decision uncertainty also are closely related. Deterministic decisions operate with little or no uncertainty because they are well understood. Non-deterministic decisions, which are influenced by what information is known at the point of decision, tend to have far more uncertainty regarding the stability or consistency of their outcomes. Figure 6 segments the decision analytics capabilities into nine categories and positions them in a framework based on the four criteria.

Figure 6 identifies nine analytic categories that support decision analytics. These categories are described as follows:

Conditional. The conditional analytic category contains algebraic expressions combining Boolean operators that express decision rules that typically take the form of “if x then y else z” or “when j then k else l.” They are highly effective at describing and automating decision processes. Conditional logic forms the basis for business rule management systems (BRMS), which can render these relationships in multiple forms (decision rules, decision tables, and decision trees). Conditional logic that is event-based provides additional support for temporal constructs of the “when j then k else l” form. Conditional logic is often combined with other analytical techniques to quantify or refine a decision, providing powerful and flexible support for decisioning.

Algorithmic. The conditional analytic category uses algebraic equations that leverage known variables and constants to create new variables. Algorithmic expressions are immensely powerful. Expressions can include transformations, reclassifications, aggregations, and functions.

Correlative. The conditional analytic category is a statistical technique that describes the strength of a relationship or dependency between variables. Simple forms of relationship analysis can include sentiment analysis or text analytics.

Optimized. Optimization is typically the maximization or minimization of an objective function subject to goals and constraints. Optimization is important because it provides a method to achieve the best possible outcome given the resources currently available.

Discrete. Discrete choice and conjoint analysis are survey-based research techniques that effectively reflect respondent preferences for a particular set of capabilities. Preferences are normalized and quantified, making them useful in understanding the relative strength of alternatives and the elasticity of demand. Survey execution also emulates the buying process, which improves data quality.

Collaborative. Collaboration is generally a more qualitative approach to decisioning, which evaluates the contributions of various constituencies including: those people who are in your circle of trust, critics, friends, and everyone else. A wide number of collaborative techniques exists. Participant contributions can be weighted; decisions can be single pass, Delphi, or stepwise; decisions can be relative or absolute; and decisions can be made by consensus, majority, plurality, committee, or autocratically.

Predictive. Predictive analytics leverages known data, relationships, and patterns to make predictions about future events. Results are sensitive to the quantity of known data and how this data is distributed.

Bayesian. Bayesian analytics enable us to understand the impact that conditional probabilities have on an outcome. Bayesian inference embraces uncertainty and develops probabilities that provide an unbiased and rational way to quantify the likelihood of an outcome or series of outcomes.

Adaptive. Adaptive systems (or complex adaptive systems) represent the frontier of decision analytics. Adaptive systems combine predictive, Bayesian analytics, economic models, and learning to govern and anticipate how to best respond to a changing environment. The challenging aspect of adaptive systems is finding new decision rules to improve operational outcomes in a changing environment while simultaneously minimizing risk.

The categories presented in the decision analytics continuum are generally mutually exclusive but selectively employed together to address decisioning.