There’s a really odd comment that I’ve sometimes heard about dealing with uncertainty. The comment goes somewhat along the lines like “oh you know, uncertainty is a problem now, but once we have good AI and algorithms our systems will be much more accurate”. I don’t think this is a very good argument for discrediting decision methods that try to grapple with uncertainty.
Why is it that we need decision methods and procedures for dealing with such situations? Why not just produce certainty and just base our decisions on that?
The annoying answer to this is that it simply costs too much. Reducing uncertainty is possible, but the more you reduce it, the more expensive it gets. To be exact, we can say that the marginal cost of uncertainty rises.
Consider an example from industrial production. Let’s say you have a production line that churns out really nice hiking boots. Unfortunately, there are some production errors every once in a while, as your line manager kindly tells you. But there is a level of uncertainty in the estimate: he is not sure how many faulty shoes will be produced in each production batch. To reduce uncertainty, you can take all kinds of measures. For example, you can hire a team of employees to inspect some of the manufactured shoes and discard any faulty ones. However, this does not eliminate uncertainty: after all, they cannot inspect every shoe. But wait, you can do more! To reduce uncertainty even more, you hire even more inspectors so that they can inspect every single shoe produced. Surely now there is no uncertainty left?
Well, unfortunately, there is. The inspectors are only human – they make errors too. So every once in a while, while one of your beloved inspectors is thinking about the upcoming football match of the evening, a faulty shoe escapes his gaze. Undaunted, you resolve to eliminate uncertainty, and fit the production line with an expensive machine inspection system. The system checks every shoe that passes the human inspectors so that they are double-checked. Surely now each produced shoe is good to go? Most days they are, until a programming error in the machine causes a problem: a shoe in an unconventional orientation is actually faulty yet passes undetected through the machine. In a fit, you eliminate the marketing department and use their funding to eliminate the uncertainty in production faults once and for all…
Why is it that we need decision methods and procedures for dealing with such situations? Why not just produce certainty and just base our decisions on that?
The annoying answer to this is that it simply costs too much. Reducing uncertainty is possible, but the more you reduce it, the more expensive it gets. To be exact, we can say that the marginal cost of uncertainty rises.
Consider an example from industrial production. Let’s say you have a production line that churns out really nice hiking boots. Unfortunately, there are some production errors every once in a while, as your line manager kindly tells you. But there is a level of uncertainty in the estimate: he is not sure how many faulty shoes will be produced in each production batch. To reduce uncertainty, you can take all kinds of measures. For example, you can hire a team of employees to inspect some of the manufactured shoes and discard any faulty ones. However, this does not eliminate uncertainty: after all, they cannot inspect every shoe. But wait, you can do more! To reduce uncertainty even more, you hire even more inspectors so that they can inspect every single shoe produced. Surely now there is no uncertainty left?
Well, unfortunately, there is. The inspectors are only human – they make errors too. So every once in a while, while one of your beloved inspectors is thinking about the upcoming football match of the evening, a faulty shoe escapes his gaze. Undaunted, you resolve to eliminate uncertainty, and fit the production line with an expensive machine inspection system. The system checks every shoe that passes the human inspectors so that they are double-checked. Surely now each produced shoe is good to go? Most days they are, until a programming error in the machine causes a problem: a shoe in an unconventional orientation is actually faulty yet passes undetected through the machine. In a fit, you eliminate the marketing department and use their funding to eliminate the uncertainty in production faults once and for all…
Uh oh, another unforeseen cause of faulty shoes!
As the example shows, reduction of uncertainty gets progressively more and more expensive every round. The more you’ve invested in it, the more you have to invest for a further reduction in uncertainty. What’s even worse – and this argument borders on the philosophical – there is practically no such thing as elimination of uncertainty. Whatever systems you come up with, there’s always a way for something really unforeseen to happen: a power failure incapacitates your inspection system, a burglar changes their settings, a meteor strikes at an inopportune time. The cause in itself is irrelevant. The point is that there’s always something you didn't anticipate.
The conclusion? There will always be some uncertainty.
And what’s more: since we have limited funds, there’s a practical limit for reducing uncertainty. At that point, we must use methods that can cope with uncertainty, because there are no other alternatives anymore.
This inevitability of facing uncertainty is why we need decision makers equipped with proper methods. Decades of behavioral decision research show (more about this in later posts) that humans are really not very good intuitive statisticians. Once you have many variables with various levels of uncertainty, there’s practically no way to make good decisions based on gut and intuition alone. What we need is methods and frameworks that simplify and aggregate information – but then again, not by too much – which we can then feed to the decision maker for processing.
The conclusion? There will always be some uncertainty.
And what’s more: since we have limited funds, there’s a practical limit for reducing uncertainty. At that point, we must use methods that can cope with uncertainty, because there are no other alternatives anymore.
This inevitability of facing uncertainty is why we need decision makers equipped with proper methods. Decades of behavioral decision research show (more about this in later posts) that humans are really not very good intuitive statisticians. Once you have many variables with various levels of uncertainty, there’s practically no way to make good decisions based on gut and intuition alone. What we need is methods and frameworks that simplify and aggregate information – but then again, not by too much – which we can then feed to the decision maker for processing.
RSS Feed
