Via Anders Sandberg

MIT is using genetic algorithms to optimise the TCP protocol:

• http://web.mit.edu/remy/
• http://web.mit.edu/remy/TCPexMachina.pdf

It seems they are trying to achieve Hofstadterian superrationality:

Why “Remy”?

In his Metamagical Themas columns in Scientific American, Douglas Hofstadter laid out a decision-theoretic framework of "superrationality." Traditional game theory looks for “rational” solutions where each player's decision is optimal after fixing that of the other players. When such a solution converges, it is a Nash equilibrium. But a Nash equilibrium may be an inefficient or unfair one. For example, in the case of the prisoner's dilemma, both prisoners are better off defecting no matter what the other prisoner does. Mutual defection is therefore the Nash equilibrium, even though each player would benefit if both cooperated instead.

Superrationality is an alternate framework where instead of fixing the decision of just the other players, each player assumes that the other players are as rational as he is, and those players make the same assumption in turn, etc. Because all players are superrational, if their positions are identical, each player can infer that the others will settle on the same strategy. The question for each player is: given that we are all going to reason similarly and end up with the same strategy, what should that strategy be?

Remy solves the problem of finding the superrational TCP for a given (uncertain or randomly-drawn) network, where all endpoints are assumed to run the same algorithm, and the question is simply what that algorithm should be to maximize each endpoint's expected value of the objective function.

In writing a program to search for a superrational TCP, we called the system Remy—the main character from the movie Ratatouille.