Why one-box?

I have sym­pa­thy with both one-box­ers and two-box­ers in New­comb’s prob­lem. Con­trary to this, how­ever, many peo­ple on Less Wrong seem to be staunch and con­fi­dent one-box­ers. So I’m turn­ing to you guys to ask for help figur­ing out whether I should be a staunch one-boxer too. Below is an imag­i­nary di­alogue set­ting out my un­der­stand­ing of the ar­gu­ments nor­mally ad­vanced on LW for one-box­ing and I was hop­ing to get help filling in the de­tails and ex­tend­ing this ar­gu­ment so that I (and any­one else who is un­cer­tain about the is­sue) can de­velop an un­der­stand­ing of the strongest ar­gu­ments for one-box­ing.

One-boxer: You should one-box be­cause one-box­ing wins (that is, a per­son that one-boxes ends up bet­ter off than a per­son that two-boxes). Not only does it seem clear that ra­tio­nal­ity should be about win­ning gen­er­ally (that a ra­tio­nal agent should not be sys­tem­at­i­cally out­performed by ir­ra­tional agents) but New­comb’s prob­lem is nor­mally dis­cussed within the con­text of in­stru­men­tal ra­tio­nal­ity, which ev­ery­one agrees is about win­ning.

Me: I get that and that’s one of the main rea­sons I’m sym­pa­thetic to the one-box­ing view but the two-box­ers has a re­sponse to these con­cerns. The two-boxer agrees that ra­tio­nal­ity is about win­ning and they agree that win­ning means end­ing up with the most util­ity. The two-boxer should also agree that the ra­tio­nal de­ci­sion the­ory to fol­low is one that will one-box on all fu­ture New­comb’s prob­lems (those where the pre­dic­tion has not yet oc­curred) and can also agree that the best time­less agent type is a one-box­ing type. How­ever, the two-boxer also claims that two-box­ing is the ra­tio­nal de­ci­sion.

O: Sure, but why think they’re right? After all, two-box­ers don’t win.

M: Okay, those with a two-box­ing agent type don’t win but the two-boxer isn’t talk­ing about agent types. They’re talk­ing about de­ci­sions. So they are in­ter­ested in what as­pects of the agent’s win­ning can be at­tributed to their de­ci­sion and they say that we can at­tribute the agent’s win­ning to their de­ci­sion if this is caused by their de­ci­sion. This strikes me as quite a rea­son­able way to ap­por­tion the credit for var­i­ous parts of the win­ning. (Of course, it could be said that the two-boxer is right but they are play­ing a pointless game and should in­stead be in­ter­ested in win­ning sim­plic­iter rather than win­ning de­ci­sions. If this is the claim then the ar­gu­ment is dis­solved and there is no dis­agree­ment. But I take it this is not the claim).

O: But this is a strange con­voluted defi­ni­tion of win­ning. The agent ends up worse off than one-box­ing agents so it must be a con­voluted defi­ni­tion of win­ning that says that two-box­ing is the win­ning de­ci­sion.

M: Hmm, maybe… But I’m wor­ried that rele­vant dis­tinc­tions aren’t be­ing made here (you’ve started talk­ing about win­ning agents rather than win­ning de­ci­sions). The two-boxer re­lies on the same defi­ni­tion of win­ning as you and so agrees that the one-box­ing agent is the win­ning agent. They just dis­agree about how to at­tribute win­ning to the agent’s de­ci­sions (rather than to other fea­tures of the agent). And their way of do­ing this strikes me as quite a nat­u­ral one. We credit the de­ci­sion with the win­ning that it causes. Is this the source of my un­will­ing­ness to jump fully on board with your pro­gram? Do we sim­ply dis­agree about the plau­si­bil­ity of this way of at­tribut­ing win­ning to de­ci­sions?

Meta-com­ment (a): I don’t know what to say here? Is this what’s go­ing on? Do peo­ple just in­tu­itively feel that this is a crazy way to at­tribute win­ning to de­ci­sions? If so, can any­one sug­gest why I should adopt the one-boxer per­spec­tive on this?

O: But then the two-boxer has to rely on the claim that New­comb’s prob­lem is “un­fair” to ex­plain why the two-box­ing agent doesn’t win. It seems ab­surd to say that a sce­nario like New­comb’s prob­lem is un­fair.

M: Well, the two-box­ing agent means some­thing very par­tic­u­lar by “un­fair”. They sim­ply mean that in this case the win­ning agent doesn’t cor­re­spond to the win­ning de­ci­sion. Fur­ther, they can ex­plain why this is the case with­out say­ing any­thing that strikes me as crazy. They sim­ply say that New­comb’s prob­lem is a case where the agent’s win­nings can’t en­tirely be at­tributed to the agent’s de­ci­sion (ig­nor­ing a con­stant value). But if some­thing else (the agent’s type at time of pre­dic­tion) also in­fluences the agent’s win­ning in this case, why should it be a sur­prise that the win­ning agent and the win­ning de­ci­sion come apart? I’m not say­ing the two-boxer is right here but they don’t seem to me to be ob­vi­ously wrong ei­ther...

Meta-com­ment (b): In­ter­ested to know what re­sponse should be given here.

O: Okay, let’s try some­thing else. The two-boxer fo­cuses only on causal con­se­quences but in do­ing so they sim­ply ig­nore all the log­i­cal non-causal con­se­quences of their de­ci­sion al­gorithm out­putting a cer­tain de­ci­sion. This is an ad hoc, un­mo­ti­vated re­stric­tion.

M: Ah hoc? I’m not sure I see why. Think about the prob­lem with ev­i­den­tial de­ci­sion the­ory. The pro­po­nent of EDT could say a similar thing (that the pro­po­nent of two-box­ing ig­nores all the ev­i­den­tial im­pli­ca­tions of their de­ci­sion). The two-boxer will re­spond that these im­pli­ca­tions just are not rele­vant to de­ci­sion mak­ing. When we make de­ci­sions we are try­ing to bring about the best re­sults, not get ev­i­dence for these re­sults. Equally, they might say, we are try­ing to bring about the best re­sults, not de­rive the best re­sults in our log­i­cal calcu­la­tions. Now I don’t know what to make of the point/​counter-point here but it doesn’t seem to me that the one-box­ing view is ob­vi­ously cor­rect here and I’m wor­ried that we’re again go­ing to end up just trad­ing in­tu­itions (and I can see the force of both in­tu­itions here).

Meta-com­ment: Again, I would love to know whether I’ve un­der­stood this ar­gu­ment and whether some­thing can be said to con­vince me that the one-box­ing view is the clear cut win­ner here.

End com­ments: That’s my un­der­stand­ing of the pri­mary ar­gu­ment ad­vanced for one-box­ing on LW. Are there other core ar­gu­ments? How can these ar­gu­ments be im­proved and ex­tended?