Re: Status of Superrationality (was Left/Right... can't we do better than this?)

From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Tue May 20 2003 - 13:50:46 MDT

  • Next message: Mike Lorrey: "Re: [GENOME] Ecce Homo?"

    Hal Finney wrote:
    >
    > A final point; Eliezer points out that the "irrational" action of
    > cooperating leads to a greater payoff. I believe the consensus
    > among game theorists is that this does not change the fact that it
    > is irrational. The reasoning is similar to that in Newcomb's paradox
    > (take one box or two, but God has arranged that you will get more money
    > if you take only one box). Taking two boxes physically entails getting
    > at least as much as in the first box, hence it is the rational action.
    > In this paradox, as well, being irrational leads to a greater outcome.
    > I copied out some analysis on this issue from a decision theory book at
    > http://www.finney.org/~hal/Newcomb.html. The argument doesn't go over
    > directly to the PD case, but the flavor is the same: it is possible for
    > an irrational action to lead to a greater outcome.

    Yes, Newcomb's Paradox is a good example of a situation which is very
    straightforward to rationally resolve for maximum benefit using the Golden
    Law(*), as discussed on the AGI list. Game theory has not caught up with
    this yet, but historically it has often taken game theorists much too long
    to realize that the "irrational" action of cooperating under situation
    XYZ, which does in fact deliver a higher payoff, is really rational after
    all. In this case solving the problem requires a timeless formulation of
    decision theory, of which ordinary decision theory is a special case. Be
    it noted for the record that Eliezer Yudkowsky disagrees with the
    consensus of game theorists about what, mathematically speaking,
    constitutes "rationality", not just in the case of the Prisoner's Dilemna,
    but also for Newcomb's Paradox and a wide variety of other situations in
    which similar or identical decision processes are distantly instantiated.
      Be it also noted that the actions Eliezer Yudkowsky computes as formally
    rational are the ones that any sane non-game-theorist would take and that
    do in fact correlate with maximum payoffs.

    (*) Not to be confused with the prescriptive Golden Rule in human
    morality, the Golden Law states descriptively that identical or
    sufficiently similar decision processes, distantly instantiated, return
    identical or correlated outputs. The prescriptive formulation of the
    Golden Law states that you should make any decision between alternatives
    as if your decision controlled the Platonic output of the mathematical
    object representing your decision system, rather than acting as if your
    decision controlled your physical instantiation alone. Wei Dai's
    formulation is even simpler; he says that you should choose between
    alternatives A and B by evaluating your utility function on the state of
    the multiverse given that your choice is A, versus the state of the
    multiverse given that your choice is B.

    But again, see the discussion on the AGI list. (Google on AGI + "Golden
    Law".)

    -- 
    Eliezer S. Yudkowsky                          http://singinst.org/
    Research Fellow, Singularity Institute for Artificial Intelligence
    


    This archive was generated by hypermail 2.1.5 : Tue May 20 2003 - 14:03:17 MDT