"Extortionate" strategy beats tit-for-tat in iterated Prisoner's Dilemma

teageegeepea

Less Wrong had a Prisoner's Dilemma contest some time back, whose results I've forgotten. Perhaps it should be rerun with William H. Press & Freeman Dyson's proposed extortionate strategies.

I hope Pinker gives a response at Edge.org, since P.D played a significant role his book "The Better Angels of Our Nature" as a source of morality embedded in the nature of logic/reality.

Hat-tip to Marginal Revolution.

Less Wrong had a Prisoner's Dilemma contest some time back, whose results I've forgotten. Perhaps it should be rerun with William H. Press & Freeman Dyson's proposed extortionate strategies.

I hope Pinker gives a response at Edge.org, since P.D played a significant role his book "The Better Angels of Our Nature" as a source of morality embedded in the nature of logic/reality.

Hat-tip to Marginal Revolution.

Perhaps it should be rerun

It is, in fact. I'm finalising the code and the announcement post. Expect it in a couple of days.

William H. Press & Freeman Dyson's proposed extortionate strategies.

As I understand them (which is admittedly not much; the paper is surprisingly dense and metaphorical) can take advantage of 'evolutionary' strategies but not strategies with a 'theory of mind' - and it appears that tit-for-tat falls under the 'theory of mind' category. Although I would state once more that I'm not sure I understand the paper.

How's your progress? You said you were "finalising the code and the announcement post", so I'm wondering whether you ran into any unexpected problems since then?

1Username14y

Fantastic! I'm very much interested in trying my hand at developing a strategy, thank you for putting this work together and I'm sure you will get a lot of positive feedback, once you finish it and release it.

0Vaniver14y

Press and Dyson's setup had two areas where 'strategies' come into play. The first area is the set of four probabilities you provide to the game, which determine your score when combined with the other player's set of four probabilities. Tit-for-tat is one particular choice of four probabilities (and, based on the nature of the game, should actually be represented as "slightly forgiving tit-for-tat", which cooperates with probability epsilon in the defect-defect or defect-cooperate case, so that when playing against itself all states will terminate with cooperate-cooperate). The second area is how the players modify those sets over time. Here, 'theory of mind' is relevant: both players with and without theories of mind can play any particular 4-probability set, like tit-for-tat. Players with theory of mind think (at least) two steps ahead- when I change my probabilities, how will my opponent change their probabilities? Players without theory of mind think only one step ahead- given my opponent's probabilities, which play maximizes my score?