The most interesting part of this talk was when he showed how people's decisions became more rational as they distanced themselves from the consequences. I thought, however, that much of his actual theory was far too simplified to be reasonable. For instance, his expectation equation is very simple: Utility * Probability of Payoff. This does not take into account possible negative payoff in the case of failure. For example, say you were going to rob a bank, which would give you a 5 million dollar payoff. Let's for now assume you do not have any moral qualms regarding robbing a bank (i.e. you're Ben Bernanke). Now let's say your probability of success is 0.05. This means your expected payoff is 5,000,000 * 0.05 = 250,000. If it costs less then 250,000 to attempt this caper, you should go for it, right? Not necessarily. What about if you fail? There is also a cost for failing. So, in every decision there is the payoff for success, there is a payoff for failure, there is a fixed price of attempting, and there is a price for not attempting. Imagine a binary decision---to act, or not to act. You could quantify the expectation of Act as:

({probability of success} * {utility of success}) +

((1 - {probability of success}) * {utility of failure}) -

{cost of attempt}.

This must be weighed against the expectation of not acting. Whichever has the higher expected value should be taken in a purely rational agent.

Another interesting point in the talk was that people make decisions by comparing the gradient. That is people prefer slopes that are moving toward higher payoffs than lower, even when the total payoff is lower. There were at least two examples of this in the talk, the salary and the burger. In the case of the salary, most people chose the salary that was increasing, even though the total net was less than the total of the salary that was decreasing. In the case of the burger, people associated its worth to what they had paid for it in the past. I do not see this as being dumb or ignorant. In fact, we have learned in nature that things at rest tend to stay at rest and things in motion tend to stay in motion. Thus, we use the slope of the payoff the predict future rewards.

These shortcuts are very interesting to me, as we may need them to create machines that are able to make good decisions---or at least machines that understand the decisions that humans are making.

## No comments:

Post a Comment