On self-experimentation in spaced repetition

I love Anki but am sufficiently devoted to spaced repetition to want more control over my training. I now spend over half my memory-training time in a custom tool.

This tool uses a much simpler scheduling algorithm: in the case of a correct answer after an interval of length L, the question comes up for review e * L later, where e is Euler's constant. So, if I answer a prompt correctly 5 days after I last saw it, it will come due in roughly 13.6 days. (Except if L is very short--there's a minimum interval of one day.)

This is a very primitive system, and I'm not interested in arguing that it's better than any existing algorithm. It has these virtues, though:

  1. It's simple enough to have implemented and tested fairly quickly. Minor and even moderate flaws in the system are acceptable, but major flaws are not.
  2. It's working surprisingly well. The reviews it schedules tend to be tougher than the ones Anki suggests on a given day.
  3. Its flaws should be easier to fix in a natural way. I have often overlooked this feature of simpler systems. Ockham's razor is often presented as a guide to truth; other exhortations to simplicity are often justified in terms of cost. This is more a benefit of natural diagnosis.
  4. It provides a benchmark against which more complicated algorithms can be tested. (This is a useful feature I associate with Marcel. If any baseball analysts are out there looking for a project, I think there's some relatively low-hanging fruit available in optimizing the various constants that appear in the Marcel system.)

I suspect that the best systems are likely to use machine learning in some form. It took me a long time to come to believe this: existing algorithms have their roots in, for example, systems for moving physical cards between boxes. These algorithms make a certain kind of sense. Their features are each intuitively motivated, and there are enough of them to make you think that the real-world phenomenon is being captured sufficiently well.

But we have computers now. The way to work with the computer, rather than against it, is often to keep it very simple until it can be totally inscrutable.


Home page