It seems the point of all this is to study the properties of programs that “cooperate” on one-shot games with known payoffs. So, there’s no point dwelling on objections as to the practicality of any of this (how to prove that an embodied computer actually implements given source code, etc. - see Ken Thompson’s C compiler/login trojan)
I’d appreciate more precision on “If G is a prisoner’s dilemma”.
Also, I suppose there’s nothing prohibiting mixed strategies (choosing a probability distribution over your likely moves, from which the actual move is drawn) - e.g. the machine the programs run on has access to random bits.
Since I didn’t actually follow your definitions carefully enough to answer, can you tell me what would be played by two identical rational CDT, Cliquish-DT, or TDT agents (or same-clique-agents) on the game:
(2,-1) (0,0)
(0,0) (0,0)
( with (row player, column player) payoffs )
Row player will always choose row 1 (if he wants to maximize the sum of payoffs, or his own payoff). What will column player choose? Take the −1 for the good of the row player?
That is, I’m asking whether programs which self-sacrifice when faced with another member of their clique are supposed to be ‘rational’ under one of these theories.
If it weren’t a one-shot game, I’d wonder about history—if an external force assigns a particular instance only to the sacrificial role, does it keep on sacrificing? I can only ask this informally, unfortunately.
I’d appreciate more precision on “If G is a prisoner’s dilemma”.
We can read off the payoff matrix. G is a Prisoner’s Dilemma if for each player, utility(I defect, they cooperate)>utility(we both cooperate)>utility(we both defect)>utility(I cooperate, they defect).
CDT and TDT would both choose column 2 in your example. CliqueBots are dependent on ad-hoc rules, so you could program them either way. There are circumstances where UDT would play column 1 against itself; we’ll get to that in a bit.
It seems the point of all this is to study the properties of programs that “cooperate” on one-shot games with known payoffs. So, there’s no point dwelling on objections as to the practicality of any of this (how to prove that an embodied computer actually implements given source code, etc. - see Ken Thompson’s C compiler/login trojan)
I’d appreciate more precision on “If G is a prisoner’s dilemma”.
Also, I suppose there’s nothing prohibiting mixed strategies (choosing a probability distribution over your likely moves, from which the actual move is drawn) - e.g. the machine the programs run on has access to random bits.
Since I didn’t actually follow your definitions carefully enough to answer, can you tell me what would be played by two identical rational CDT, Cliquish-DT, or TDT agents (or same-clique-agents) on the game:
( with (row player, column player) payoffs )
Row player will always choose row 1 (if he wants to maximize the sum of payoffs, or his own payoff). What will column player choose? Take the −1 for the good of the row player?
That is, I’m asking whether programs which self-sacrifice when faced with another member of their clique are supposed to be ‘rational’ under one of these theories.
If it weren’t a one-shot game, I’d wonder about history—if an external force assigns a particular instance only to the sacrificial role, does it keep on sacrificing? I can only ask this informally, unfortunately.
We can read off the payoff matrix. G is a Prisoner’s Dilemma if for each player, utility(I defect, they cooperate)>utility(we both cooperate)>utility(we both defect)>utility(I cooperate, they defect).
That seems clear. Perhaps in addition, u(c,d)+u(d,c) < 2*u(c,c), though I’m not sure if that makes a difference to your claims.
Again, no difference for CDT or TDT (or the version of ADT I’ll present), but sometimes it matters for UDT.
CDT and TDT would both choose column 2 in your example. CliqueBots are dependent on ad-hoc rules, so you could program them either way. There are circumstances where UDT would play column 1 against itself; we’ll get to that in a bit.