Yeah, I know that I’m somewhat sloppily identifying the priors with the programs that make up most of their weight. It makes it more convenient for me to think about what agents using those priors would do—though I am probably missing some details that would stem from using a computable approximation rather than the possibly uncomputable one of looking at the most successful few programs.
And it serves me right for not googling “speed prior.” I forgot it was length+log(time). I’m sure computational complexity people would have much more interesting things to say than me about why that might get you plenty of mathematics programs that are neither brute force nor lookup tables. Or maybe it’s just that taking the log of time to turn it into “bits” is the number one obvious thing to do.
If we think of the Solomonoff prior as the “real answer” and a speed-like prior as a convenience used for making computable predictions when computation has a cost, we could get some kind of principled answer from estimates of the benefit of predictive accuracy and the cost of time. I spent a few minutes and couldn’t figure it out—worst case reasoning about predictive accuracy gets stuck on some very rare worst cases, and I’m not sure how to do the average case reasoning right.
Nice to meet you again too :)
Yeah, I know that I’m somewhat sloppily identifying the priors with the programs that make up most of their weight. It makes it more convenient for me to think about what agents using those priors would do—though I am probably missing some details that would stem from using a computable approximation rather than the possibly uncomputable one of looking at the most successful few programs.
And it serves me right for not googling “speed prior.” I forgot it was length+log(time). I’m sure computational complexity people would have much more interesting things to say than me about why that might get you plenty of mathematics programs that are neither brute force nor lookup tables. Or maybe it’s just that taking the log of time to turn it into “bits” is the number one obvious thing to do.
If we think of the Solomonoff prior as the “real answer” and a speed-like prior as a convenience used for making computable predictions when computation has a cost, we could get some kind of principled answer from estimates of the benefit of predictive accuracy and the cost of time. I spent a few minutes and couldn’t figure it out—worst case reasoning about predictive accuracy gets stuck on some very rare worst cases, and I’m not sure how to do the average case reasoning right.