On Designing a Decision Intelligence System for Uncertainty

Part 2: Handling Epistemic Uncertainty

Photo by Google DeepMind on Unsplash

This is a multi-part post to discuss strategies for enhancing the design of a decision intelligence system to accommodate uncertainty.

In part 1 of this series, I shared several broad observations o help establish design criteria around uncertainty. In this post, I will discuss the impact of epistemic uncertainties in the decision intelligence system design.

What is Epistemic Uncertainty?

Numerous situations exist where certain information necessary for decision-making is absent, but we know there is a way to procure the missing piece. This type of uncertainty is called Epistemic Uncertainty. The term “epistemic” is derived from the Greek word “episteme,” meaning knowledge.

Take, for instance, the decision to purchase a house. It is essential first to understand the size of the mortgage for which you’re eligible. We can approach the bank and apply for prequalification for a mortgage.

Epistemic uncertainty is more manageable and can be diminished by collecting additional information, undertaking further research, or employing more advanced modeling methods. That is why epistemic uncertainty is also known as Reducible Uncertainty.

However, it’s crucial to be aware that while mitigating epistemic uncertainty, we might inadvertently spawn other uncertainties that aren’t as easily tamed. Unconscious of this potential pitfall, we could inadvertently opt for a less effective uncertainty reduction method instead of a more beneficial one.

Impact on Decision Intelligence System Design

At first glance, it may not seem like a significant issue. One could presume that simply incorporating a component into the design to gather the missing knowledge would resolve the matter. However, it can sometimes be more complicated than imagined.

A simple-minded approach may only address the issue partially. It’s essential to delve deeper into the specifics and intricacies of the situation to develop a comprehensive and effective solution.

Accommodating epistemic uncertainty in the design of a decision intelligence system requires further investigation into the nature of epistemic uncertainty.

At the highest level, there are two scenarios of epistemic uncertainty:

1. You know what you don't know (The case of “Known Unknowns”)
2. You don't know what you don't know (The case of “Unknown Unknowns”)

Let us address these two scenarios individually.

The Case of “Known Unknowns”

If you know what you don’t know, a two-step approach is sufficient:

1. Develop a strategy and plan to gather each missing knowledge item.
2. Gather missing information by using the strategy developed in step 1.

Let us discuss what a strategy means and what is involved in developing it. The strategy here pertains to establishing a method for gathering information and deciding from where the information will be gathered. When multiple methods and sources are available for the same knowledge item, one must choose the most effective ones using one or more decision criteria. This is like an auxiliary decision-making problem of its own.

First, let us consider the nature of information sources:

1. Information exists in a data repository
2. Information exists in a person’s head
3. The information does not exist yet

The decision intelligence system design must account for all these three situations.

Data Repository as the Information Source: In this case, you must include a component module to interact with various data sources if it is not already a part of your design. Additionally, you’ll require a component that manages data sources, including the capacity to introduce new ones as necessary.

An Individual as the Information Source: The individual may be a decision-maker, stakeholder, subject-matter expert, or simply a user. In this case, it’s essential to incorporate a module that facilitates human input. This change could be as straightforward as adding a form-based interface for direct input or may involve a more complex component designed to gather human beliefs, judgments, and preferences. Additionally, a feature that manages the roster of human input providers and the schedule for capturing their input will be necessary.

The information does not exist yet: In such situations, it becomes necessary to undertake a task that generates the absent information. One may accomplish this through collecting real-world observations, forecasting information using developed models, or initiating a more comprehensive information-gathering project. Regardless of the approach, a design component should be in place to schedule, assign resources, initiate a task (or a project), and monitor its progress.

The Case of “Unknown Unknowns”

Handling uncertainties we are unaware of is one of the most challenging aspects of decision intelligence system design.

If you’re navigating a landscape of “unknown” unknowns, you will have to adopt the following approach:

1. Uncover “unknown unknowns” as much as possible upfront
2. Learn and Adapt over time

Uncovering “Unknown Unknowns”

Once you uncover an “Unknown Unknown” it becomes a “Known Unknown” case and can be handled as described earlier. Uncovering unknown unknowns is a deliberate effort undertaken early on in the design process. Here are some strategies to achieve that:

• Engage diverse perspectives: Different people might be aware of different facets of decision-making in their practice, helping uncover unknown unknowns. This would require design components for team brainstorming support. You may also include a generative AI component for research and ideation support.
• Review Decision Problem Framing: Reviewing and questioning existing decision problem framing often leads to the expansion of problem boundaries, helping uncover unknown unknowns. This would require a design module for decision problem framing. The decision problem framing module may include guidance generated from generative AI.

Learning and Adapting Over Time

Despite one’s best efforts, it is not possible to uncover all unknown unknowns upfront. The system design may choose to ignore all that is not uncovered or, alternatively, design the system to be adaptive. Here are some strategies to consider:

• Embrace Flexibility: Design the system to be flexible and adaptive so it can easily accommodate new types of data and its cascading impact on user interface and logic.
• Regular Reviews and Updates: Regularly review the system and do updates to accommodate new insights or information. Continuous learning and improvement should be vital to any decision intelligence system. This would require a design component for periodic (manual or automatic) review of the system state, a module for system monitoring and recommending changes, and a module for the change management process.

The design changes for handling unknown unknowns could necessitate a substantial system redesign. However, depending on the system’s purpose and constraints, you may forego handling this uncertainty type.

Summary of Design Impact

The following table summarizes the impact of epistemic uncertainty on the design of a decision intelligence system:

Design Components List

What’s Next?

After addressing all epistemic uncertainties, evaluating your solutions is essential to discern if they may inadvertently introduce other forms of uncertainty.

In part 3 of this series, I will delve deeper into these potential new uncertainties and explore how they influence the design of your decision intelligence system.