The Scylla of Error and the Charybdis of Paralysis
We’re interested in improving human rationality. Many of our techniques for improving human rationality take time. In real-time situations, you can lose by making the wrong decision, or by making the “right” decision too slowly. Most of us do not have inflexible-schedule, high-stakes decisions to make, though. How often does real-time decision making really come up?
Suppose you are making a fairly long-ranged decision. Call this decision 1. While analyzing decision 1, you come to a natural pause. At this pause you need to decide whether to analyze further, or to act on your best-so-far analysis. Call this decision 2. Note that decision 2 is made under tighter time pressure than decision 1. This scenario argues that decision-making is recursive, and so if there are any time bounds, then many decisions will need to be made at very tight time bounds.
A second, “covert” goal of this post is to provide a definitely-not-paradoxical problem for people to practice their Bayseian reasoning on. Here is a concrete model of real-time decisionmaking, motivated by medical-drama television shows, where the team diagnoses and treats a patient over the course of each episode. Diagnosing and treating a patient who is dying of an unknown disease is a colorful example of real-time decisionmaking.
To play this game, you need a coin, two six-sided dice, a deck of cards, and a helper to manipulate these objects. The manipulator sets up the game by flipping a coin. If heads (tails) the patient is suffering from an exotic fungus (allergy). Then the manipulator prepares a deck by removing all of the clubs (diamonds) so that the deck is a red-biased (black-biased) random-color generator. Finally, the manipulator determines the patients starting health by rolling the dice and summing them. All of this is done secretly.
Play proceeds in turns. At the beginning of each turn, the manipulator flips a coin to determine whether test results are available. If test results are available, the manipulator draws a card from the deck and reports its color. A red (black) card gives you suggestive evidence that the patient is suffering from a fungus (allergy). You choose whether to treat a fungus, allergy, or wait. If you treat correctly, the manipulator leaves the patient’s health where it is (they’re improving, but on a longer timescale). If you wait, the manipulator reduces the patient’s health by one. If you treat incorrectly, the manipulator reduces the patient’s health by two.
Play ends when you treat the patient for the same disease for six consecutive turns or when the patient reaches zero health.
Here is some Python code simulating a simplistic strategy. What Bayesian strategy yields the best results? Is there a concise description of this strategy?
The model can be made more complicated. The space of possible actions is small. There is no choice of what to investigate next. In the real world, there are likely to be diminishing returns to further tests or further analysis. There could be uncertainty about how much time pressure there is. There could be uncertainty about how much information future tests will reveal. Every complication will make the task of computing the best strategy more difficult.
We need fast approximations to rationality (even quite bad approximations, if they’re fast enough), as well as procedures that spend time in order to purchase a better result.