One Doubt About Time­less De­cision Theories

Time­less De­cisions The­or­ies (in­clud­ing vari­ants like FDT, UDT, ADT, ect.) provide a rather el­eg­ant method of solv­ing a broader class of prob­lems than CDT. While CDT re­quires the out­comes of de­cisions to be in­de­pend­ent of the in­di­vidual mak­ing the de­cision (in such a way that causal sur­gery on a single node is valid), time­less de­cisions the­or­ies can handle any prob­lem where the out­come is a func­tion of the choice se­lec­ted (even if this oc­curs in­dir­ectly as a res­ult of a pre­dic­tion).

(Epistemic Status: Thoughts for fur­ther in­vest­ig­a­tion)

This is an ex­cel­lent reason to in­vest­ig­ate these de­cision the­or­ies, yet we need to make sure that we don’t get blinded by in­sight. Be­fore we im­me­di­ately jump to con­clu­sions by tak­ing this im­prove­ment, it is worth­while con­sid­er­ing what we give up. Per­haps there are other classes that we might which to op­tim­ise over which we can no longer op­tim­ise over once we have in­cluded this whole class?

After all, there is a sense in which there is no free lunch. As dis­cussed in the TDT pa­per, for any al­gorithm, we could cre­ate a situ­ation where there is an agent that spe­cific­ally pun­ished that al­gorithm. The usual re­sponse is that these situ­ations are un­fair, but a) the uni­verse is of­ten un­fair b) there are plaus­ible situ­ations where the al­gorithm chosen in­flu­ences the out­come is slightly less un­fair ways.

Ex­pand­ing on b), there are times when you want to be pre­dict­able to sim­u­lat­ors. Indeed, I can even ima­gine agents that wish to elim­in­ate agents that they can’t pre­dict. Fur­ther, rather than fa­cing a per­fect pre­dictor, it seems like it’ll be at least a few or­ders of mag­nitude more likely that you’ll face an im­per­fect pre­dictor. Model­ling these as X% per­fect pre­dictor, 100-X% ran­dom pre­dictor will usu­ally be im­plaus­ible as pre­dict­ors won’t have a uni­form suc­cess rate over all al­gorithms. These situ­ations are slightly more plaus­ible for scen­arios in­volving AI, but even if you per­fectly know an agent’s source code, you are un­likely to know its ex­act ob­ser­va­tional state due to ran­dom noise.

It there­fore seems that the “best” de­cision the­ory al­gorithm might be dom­in­ated by factors other than op­timal per­form­ance on the nar­row class of prob­lems TDT op­er­ates on. It may very well be the case that TDT is ul­ti­mately tak­ing the right ap­proach, but even if this is the case, I thought it was worth­while sketch­ing out these con­cerns so that they can be ad­dressed.