I don’t know too much about decision theory, but I was thinking about it a bit more, and for me, the end result so far was that “dominant strategy” is just a flawed concept.
If the agents behave superrationally, they do not care about the dominant strategy, and they are safe from this attack. And the “super” in superrational is pretty misleading, because it suggests some extra-human capabilities, but in this particular case it is so easy to see through the whole ruse, one has to be pretty dumb not to behave superrationally. (That is, not to consider the fact that other agents will have to go though the same analysis as ourselves.)
Superrationality works best when we actually know that the others have the same input-output function as ourselves, for example when we know that we are clones or software copies of each others. But real life is not like that, and now I believe that the clean mathematical formulation of such dilemmas (with payoff matrices and all that) is misleading, because it sweeps under the rug another, very fuzzy, hard to formalize input variable: the things that we know about the reasoning processes of the other agents. (In the particular case of the P+epsilon attack, we don’t have to assume too much about the other agents. In general, we do.)
I don’t know too much about decision theory, but I was thinking about it a bit more, and for me, the end result so far was that “dominant strategy” is just a flawed concept.
If the agents behave superrationally, they do not care about the dominant strategy, and they are safe from this attack. And the “super” in superrational is pretty misleading, because it suggests some extra-human capabilities, but in this particular case it is so easy to see through the whole ruse, one has to be pretty dumb not to behave superrationally. (That is, not to consider the fact that other agents will have to go though the same analysis as ourselves.)
Superrationality works best when we actually know that the others have the same input-output function as ourselves, for example when we know that we are clones or software copies of each others. But real life is not like that, and now I believe that the clean mathematical formulation of such dilemmas (with payoff matrices and all that) is misleading, because it sweeps under the rug another, very fuzzy, hard to formalize input variable: the things that we know about the reasoning processes of the other agents. (In the particular case of the P+epsilon attack, we don’t have to assume too much about the other agents. In general, we do.)