Updateless decision theories


by Eliezer Yudkowsky Aug 4 2016

Decision theories that maximize their policies (mappings from sense inputs to actions), rather than using their sense inputs to update their beliefs and then selecting actions.

Updateless decision theories maximize over policies--mappings from sense inputs to actions--instead of using their sense inputs to update their beliefs, and then selecting actions using the updated model. In normal problems, these two principles of rational choice will yield identical policies; but on some Newcomblike problems they may yield different policies. Updateless decision theories refuse bad bargains in the Ultimatum Game, resist simple forms of blackmail, and are Counterfactually Muggable. Their fair problem class corresponds to "cases where my payoff depends on my behavior given any kind of sense input", where an updateful TDT would have a fair problem class only of "cases where my payoff only depended on my behavior given the sense inputs I actually got". Coinvented by Wei Dai and Vladimir Nesov, and the second major update after the invention of timeless decision theory.