Asymptotic Decision Theory

Lately I’ve been working on writing up some decision theory stuff I did with some of the other MIRI folk. The line of research was kickstarted by Jessica Taylor, and I spent a lot of time with Sam Eisenstat and Tsvi Benson-Tilsen working on it over the following month. You can find the pdf here.

UPDATE: I’ve decided to move the open questions section to here.

There are still a lot of questions I’d like answered in the realm of asymptotic decision theory.

While a lot of problems we’ve historically been interested in turn out to be convergent, there are still plenty of useful benchmarks which aren’t. For example, the problem “predict this sequence of bits” is not convergent. How would we go about behaving optimally on non-convergent problems? Relatedly, is there a crisper characterization of convergence in the sense that continuity was a cripser characterization of fairness?

I’ve also been looking into the behavior of sadt and ddt in game theory problems, which I’ll be talking more about in my next post. Annoyingly enough, it turns out that most games against sadt and ddt are unfair problems!