The idea is to compare not the results of actions, but the results of decision algorithms. The question that the agent should ask itself is thus:

"Suppose everyone1 who runs the same thinking procedure like me uses decision algorithm X. What utility would I get at the 50th percentile (not: what expected utility should I get), after my life is finished?"
Then, he should of course look for the X that maximizes this value.

Now, if you formulate a turing-complete "decision algorithm", this heads into an infinite loop. But suppose that "decision algorithm" is defined as a huge table for lots of different possible situations, and the appropriate outputs.

Let's see what results such a thing should give:

  • If the agent has the possibility to play a gamble, and the probabilities involved are not small, and he expects to be allowed to play many gambles like this in the future, he should decide exactly as if he was maximizing expected utility: If he has made many decisions like this, he will get a positive utility difference in the 50th percentile if and only if his expected utility from playing the gamble is positive.
  • However, if Pascal's mugger comes along, he will decline: The complete probability of living in a universe where people like this mugger ought to be taken seriously is small. In the probability distribution over expected utility at the end of the agent's lifetime, the possibility of getting tortured will manifest itself only very slightly at the 50th percentile - much less than the possibility of losing 5 Dollars.

The reason why humans will intuitively decline to give money to the mugger might be similar: They imagine not the expected utility with both decisions, but the typical outcome of giving the mugger some money, versus declining to.

1I say this to make agents of the same type cooperate in prisoner-like dilemmas.

New to LessWrong?

New Comment
20 comments, sorted by Click to highlight new comments since: Today at 3:06 PM

What happens if you're using this method and you're offered a gamble where you have a 49% chance of gaining 1000000utils and a 51% chance of losing 5utils (if you don't take the deal you gain and lose nothing). Isn't the "typical outcome" here a loss, even though we might really really want to take the gamble? Or have I misunderstood what you propose?

Depending on the rest of your utility distribution, that is probably true. Note, however, that an additional 10^6 utility in the right half of the utility function will change the median outcome of your "life": If 10^6 is larger than all the other utility you could ever receive, and you add a 49 % chance of receiving it, the 50th percentile utility after that should look like the 98th percentile utility before.

Could you rephrase this somehow? I'm not understanding it. If you actually won the bet and got the extra utility, your median expected utility would be higher, but you wouldn't take the bet, because your median expected utility is lower if you do.

In such a case, the median outcome of all agents will be improved if every agent with the option to do so takes that offer, even if they are assured that it is a once/lifetime offer (because presumably there is variance of more than 5 utils between agents).

But the median outcome is losing 5 utils?

Edit: Oh, wait! You mean the median total utility after some other stuff happens (with a variance of more than 5 utils)?

Suppose we have 200 agents, 100 of which start with 10 utils, the rest with 0. After taking this offer, we have 51 with -5, 51 with 5, 49 with 10000, and 49 with 10010. The median outcome would be a loss of -5 for half the agents, a gain of 5 for half, but only the half that would lose could actually get that outcome...

And what do you mean by "the possibility of getting tortured will manifest itself only very slightly at the 50th percentile"? I thought you were restricting yourself to median outcomes, not distributions? How do you determine the median distribution?

And what do you mean by "the possibility of getting tortured will manifest itself only very slightly at the 50th percentile"? I thought you were restricting yourself to median outcomes, not distributions? How do you determine the median distribution?

I don't. I didn't write that.

Your formulation requires that there be a single, high probability event that contributes most of the utility an agent has the opportunity to get over its lifespan. In situations where this is not the case (e.g. real life), the decision agent in question would choose to take all opportunities like that.

The closest real-world analogy I can draw to this is the decision of whether or not to start a business. If you fail (which there is a slightly more than 50% chance you will), you are likely to be in debt for quite some time. If you succeed, you will be very rich. This is not quite a perfect analogy, because you will have more than one chance in your life to start a business, and the outcomes of business ownership are not orders of magnitude larger than the outcomes in real life. However, it is much closer than the "51% chance to lose $5, 49% chance to win $10000" that your example intuitively brings to mind.

Ah! Sorry for the mixed-up identities. Likewise, I didn't come up with that "51% chance to lose $5, 49% chance to win $10000" example.

But, ah, are you retracting your prior claim about a variance of greater than 5? Clearly this system doesn't work on its own, though it still looks like we don't know A) how decisions are made using it or B) under what conditions it works. Or in fact C) why this is a good idea.

Certainly for some distributions of utility, if the agent knows the distribution of utility across many agents, it won't make the wrong decision on that particular example by following this algorithm. I need more than that to be convinced!

For instance, it looks like it'll make the wrong decision on questions like "I can choose to 1) die here quietly, or 2) go get help, which has a 1/3 chance of saving my life but will be a little uncomfortable." The utility of surviving presumably swamps the rest of the utility function, right?

Ah, it appears that I'm mixing up identities as well. Apologies.

Yes, I retract the "variance greater than 5". I think it would have to be variance of at least 10,000 for this method to work properly. I do suspect that this method is similar to decision-making processes real humans use (optimizing the median outcome of their lives), but when you have one or two very important decisions instead of many routine decisions, methods that work for many small decisions don't work so well.

If, instead of optimizing for the median outcome, you optimized for the average of outcomes within 3 standard deviations of the median, I suspect you would come up with a decision outcome quite close to what people actually use (ignoring very small chances of very high risk or reward).

This all seems very sensible and plausible!

A bounded utility function, on which increasing years of happy life (or money, or whatever) give only finite utility in the infinite limit, does not favor taking vanishing probabilities of immense payoffs. It also preserves normal expected utility calculations so that you can think about 90th percentile and 10th percentile, and lets you prefer higher payoffs in probable cases.

Basically, this "median outcome" heuristic looks like just a lossy compression of a bounded utility function's choice outputs, subject to new objections like APMason's. Why not just go with the bounded utility function?

I want that it is possible to have a very bad outcome: If I can play a lottery that has 1 utilium cost, 10^7 payoff and a winning chance of 10^-6, and if I can play this lottery enough times, I want to play it.

"Enough times" to make it >50% likely that you will win, yes? Why is this the correct cutoff point?