The best kittens, technology, and video games blog in the world.

Friday, October 18, 2013

Akrasia Theory and Beeminder update

risky game by mathias-erhart from flickr (CC-SA)

So this post will be about two only loosely related things - some problems with akrasia theory and my Beeminder status update.

Akrasia Theory

First, akrasia theory. The theoretical underpinning of it is that humans and animals for that matter use "incorrect" and time-inconsistent hyperbolic discounting instead of much more "correct" and time-consistent exponential discounting.

This is extremely problematic from evolutionary point of view - since any animal that moves even one bit closer towards "correct" discounting models could be massively more successful than ones that use "incorrect" models.

There's a bunch of ways to deal with this issue. A hopefully exhaustive list would be:
  1. Exponential discounting is more correct, but brains cannot implement exponential discounting even in principle
  2. Exponential discounting is more correct, brains can implement exponential discounting in principle, and it would be major evolutionary advantage, but in never evolved yet
  3. Exponential discounting is more correct, brains can implement exponential discounting in principle, but it wouldn't bring major evolutionary advantage
  4. Hyperbolic discounting is more correct
The first of these is just extremely problematic. Bounded rationality is fine, but exponential discounting is mathematically extremely simple, and it's hard to come up with any argument why it would be even tiny bit more complicated or difficult to implement than hyperbolic discounting - something even tiniest animal brains seem to have no problems with. Hyperbolic discounting has a small advantage that natural scale for most things is logarithmic, not linear, but we can think about linear scales if we want to. What's worse - if we accept that linear scales are impossible to think with, that completely destroys any chance of game theory or any other mathematical theory of mind working - and nearly 100% of things evolutionary psychology postulates are far more complex than that, so we'd have to reject them all a priori (I'm quite willing to reject most of them for this reason anyway, but not to such extreme).

The second alternative is actually somewhat promising, since we know of huge number of extremely successful evolutionary traits that evolved only once, even though nothing obvious prevented them from evolving much earlier. Like basically everything humans do. Even simple things like ability to synthesize all essential vitamins is entirely possible in principle, and it would be advantageous to organisms that lost it (which is all animals, pretty much) - but once organisms lost such ability, they never regain it. Still, it would really deserve some explanation why this is a particularly difficult thing to evolve into, and it's hard to come up with any.

The third alternative is probably the most popular (just after pretending the problem doesn't exist) - sure, exponential discounting would be better, but then animals' decision processes are based on duct taping huge number of silly heuristics together, so whatever problems hyperbolic discounting causes have been papered over by other biases and heuristics anyway, so it's not a big deal outside completely artificial laboratory conditions. It's difficult to come up with a convincing counterargument against this. It forces us to abandon all mathematically simple of human or animal behaviour, but these were totally silly in the first place anyway.

The last alternative is something that almost nobody except me seems to take seriously, and perhaps they're all right, but maybe - just maybe - animals use hyperbolic discounting because it models the real world better than exponential discounting? The thing is - exponential processes despite their mathematical simplicity are nowhere to be found in the nature. Even radioactive decay - the classic example of exponential process - looks much more hyperbolic once you move from a single element to a mix of elements (or include secondary products of radioactive decay in the distribution). I have no way to prove it, but it definitely feels right, and avoids completely ridiculous conclusions that come from taking exponential discounting seriously.

sleeps with bees by splityarn from flickr (CC-NC-SA)

Beeminder Update

Anyway, here's my regular Beeminder update. My 6 "do more" goals, even after previous ramp up got to average of 34 days until fail, which basically means Beeminder is currently completely toothless. So far it looks like I'm not going to pay them a single dollar ever, but then who knows.

I feel it's important to distinguish between target levels (I wish to do X/week) and commitment levels (I commit to do X/week). Commitment should be much lower than target. By how much depends on variance - for low-variance goals like exercise they can be pretty close, for high-variance goals you should commit to much less than you target, at least initially. If it works, you can always increase commitment levels.

Goal status, from hardest to easiest:
  • Exercise. It's going most smoothly of all goals. I'm going to increase it from 2.5h/week to 3h/week commitment (current average is 3.4h/week). All this is in wii fit / ddr "minutes", and 3h of such "minutes" is actually about 3h45m wall clock time. I have no plans to ever commit myself to anything more than that, since there's only so much time in the week and there are diminishing health benefits to exercise. If that's ever too easy I'll just increase exercise intensity instead.
  • Online Education. I found a bunch of alternative resources since then, so it's not just Udacity now. The problem is conversion factors between them, and I'm taking a fairly generous interpretation here that full unit is full unit, even if most are much easier than Udacity's GPGPU course. I'm increasing this from 2.5/week to 3/week, and I might increase this again in the future, depending on what other education sources I'll find.
  • Play Magic. I could easily increase the commitment to play 10 games/week, but that's a good number and there's no reason to overdo it. I'm definitely playing enough to stay up to date with all the developments in the game, and I'll explore other formats some more once I get bored with Theros Standard.
  • Open Source. This is going really well right now. I feel this is extremely important, so I'm just increasing it from 3 commits/week to 7 commits/week. This might increase further, depending on circumstances.
  • Blog. My blogging has been really active recently, and I plan to keep it this way, so I'm doubling the rate from 1 posts/week to 2 posts/week. This also might increase a lot more.
  • Books. There's a big problem of what counts as a "book" here, since some are a lot longer than others. I'll be going through a ton of audiobooks for exercise anyway, and still I have a ton of paper and Kindle books to read in addition to that, so I'll increase it from 1 book/3 weeks to 1 book/2 weeks to maybe get through all these stacks someday. Perhaps I'll change this goal to count shorter books for less and longer books for more someday? We'll see.
I also have a bunch of things I want to measure but I don't yet feel like committing to them on beeminder. I added one of them - a "do less" goal to give 80% paleo diet a try, and so far it's going quite well. The biggest position on the non-paleo budget is unsurprisingly coffee, which just isn't particularly good without sugar, milk, and chocolate, and considering my coffee requirements and realistic off-paleo days for external reasons I doubt I'll be tightening that 20% non-paleo allowance anytime soon. To be honest I haven't noticed any interesting effects from that, other than it forced me to try different things in the kitchen.

As for my other non-beeminder measures, two of them are reasonably on-track, and one is very much off-track. No idea if I'm going to commit to them or not anytime soon, but I'm thinking about it.


dreeves said...

Highly astute question! I've always treated it as something that just worked in the ancestral environment -- like the heuristic of stuffing your face with as much fat and sugar as you can find -- but not in the modern environment.

Probably relevant to this question is Ainslie's study of akrasia in pigeons:

taw said...

dreeves: OK, so I checked the paper. If I understand it correctly, it's basically saying pigeons are willing to get 3x less food if only it comes 3s earlier. That's not hyperbolic discounting - that's just ridiculous (in addition to ton of other issues the paper mentions that imply very low reliability of the result).

I feel all of that is that's just an artifact of experimental design, and the effect is due to something completely different.

Exponential discounting implies a lot of crazy things - I feel they're much worse than hyperbolic discounting's dynamic inconsistency. I'll elaborate on that in some future post.