There’s a class of mimicking players that I’d like to discuss. Pseudocode of an example:
def mimic_bot(opponent):
if random() > ε:
my_source = quine_my_source()
return eval(opponent, my_source) # do as your opponent does to you
else:
# unlikely, but necessary to break out of infinitely recursive simulations
return COOPERATE
MimicBot’s strategy will almost perfectly mirror its opponent’s. Is there literature on a MimicBot, MirrorBot, PsychicTitForTatBot or something like that?
Aside: The source for MimicBot does not contain the DEFECT symbol, which might make it easier† for other programs to prove that MimicBot won’t defect against them.
This algorithm is now published in “Robust program equilibrium” by Caspar Oesterheld, Theory and Decision (2019) 86:143–159, https://doi.org/10.1007/s11238-018-9679-3, which calls it ϵGroundedFairBot.
The paper cites this comment by Jessica Taylor, which has the version that uses reflective oracles (NicerBot). Note also the post by Stuart Armstrong it’s responding to, and the reply by Vanessa Kosoy. The paper also cites a private conversation with Abram Demski. But as far as I know, the parent to this comment is older than all of these.
There’s a class of mimicking players that I’d like to discuss. Pseudocode of an example:
MimicBot’s strategy will almost perfectly mirror its opponent’s. Is there literature on a MimicBot, MirrorBot, PsychicTitForTatBot or something like that?
Aside: The source for MimicBot does not contain the DEFECT symbol, which might make it easier† for other programs to prove that MimicBot won’t defect against them.
† easier but not trivial
This algorithm is now published in “Robust program equilibrium” by Caspar Oesterheld, Theory and Decision (2019) 86:143–159, https://doi.org/10.1007/s11238-018-9679-3, which calls it ϵGroundedFairBot.
The paper cites this comment by Jessica Taylor, which has the version that uses reflective oracles (NicerBot). Note also the post by Stuart Armstrong it’s responding to, and the reply by Vanessa Kosoy. The paper also cites a private conversation with Abram Demski. But as far as I know, the parent to this comment is older than all of these.
AFAIK the existing literature on program equilibrium has not dealt with random or even pseudorandom strategies.