Problem Selection & Definition
Handbook

Selecting the Right Problem

Most of us often ask, “What is AI even capable of?”, “What can I do with AI to solve a particular problem?” Instead, the key is to remove the term ‘AI’ from the question to eliminate misconceptions. Instead, define the problem in terms of data and experience. Colloquially, Data refers to external information required to make a judgment. Experience refers to past success or failure. A typical brainstorming session should at least touch upon these two topics: what data would someone need to make a certain decision, and what experience would they likely need in order to be successful? These are not trivial, ‘introductory’ questions—most real discussions of the possibility of AI to solve a problem stem from them. The words ‘data’ and ‘experience’ are actually stand-ins for similar concepts in machine learning, and therefore translate well to the technical domain.

More on Data and Experience

In many systems, data and experience are both represented as the same kind of information. In other systems, data and expereince are quite different. For example, in a food reviewing app, data may constitute information about each dish, while experience would represent ratings. In a predictive scenario, data may exactly equal past experience.

Sometimes, you already have a certain piece of data or a certain repository of experience, but you are looking for possible AI solutions that may benefit from it. In that case, you can use the same questions phrased conversely: what decisions could be made given these data-points, and what situations may benefit from this experience?

An example may make this process of problem selection more clear. Supposing a person wanted to estimate housing prices in a fluctuating market. This estimation would likely require some input information—address, number of bedrooms and bathrooms, material, year of construction, and other similar details. Crucially, many of these data points are not strictly necessary, but may increase the certainty of a given prediction. However, is this data actually enough? Answering this is straightforward: just give a list of such data to a person, with the house price blacked out on a few entries, and ask the person to estimate the blacked out entries. In fact, going through such an exercise may reveal some elements of the problem that were not obvious initially. Perhaps the name of the city where the house is located matters quite a bit more than the exact number of bathrooms, etc. Next, consider what it means for a person who estimates house prices to be experienced—does this mean they adjust their estimations based on past feedback? If so, then an AI system may also require such a capability.

Posing these questions to your team allows each person’s expertise to shine through. Perhaps one person with a legal background can share their experiences with real-estate disputes. Perhaps a policy-minded individual may give a flavor as to what forms of data carry risks and regulations. This way, even if the discussion enters a more technical realm, non-technical individuals should feel comfortable reasoning about the implications of the downstream solutions.

By working through this exercise of data and experience, a great many ideas can be found, triaged, evaluated, and even tested for feasibility. Even if you can’t confirm that such an idea will translate into a successful AI, you can at least disconfirm it if the idea fails to pass rigor at this stage.