A follow-up to Decision theories, LW-style.

The contradiction goes like this: you want your players to use a very specific reasoning process that you like (“I will do whatever is better for me”), but somehow end up cooperating.

Let’s say I propose the following algorithm: “always cooperate”. You go:

But let’s consider the first player. Isn’t it better for them to defect?

I say “sure, but they don’t do what is better for them”. You retort:

Then the payoff matrix is different!

I say “the payoff matrix is the same, the algorithm is different”. You beg to disagree, and proceed to construct a payoff matrix that corresponds to my algorithm: “cooperate = 1, defect = 0”. I comment: “sure, IF YOU INSIST ON SELFISH PLAYERS, then you would have to use a different payoff matrix to replicate the behavior of cooperating players, and now your game is not a prisoner’s dilemma anymore”.

Here is an uncharitable interpretation. The dilemma does not arise when you explicitly treat the players as dumb automatons that you want to cooperate; it only arises when you put yourself into the shoes of one of the players. “I want to cooperate, but I want to choose it on my own and not be forced into it.”

Stop it! If you only consider decision algorithms that are acceptable for you to adopt, you will never find a good one. You will never consider, for instance, stealing tricks from the religion’s book to create better-cooperating players.

Decision algorithms are not for you. They are for stupid agents without free will. You have the power to make them do whatever you want. Don’t try to bestow free will upon them – they don’t need it. And don’t limit them to what you would do.