7 ms·
Yes. Bandits will often converge more quickly to the optimal strategy, but it is much more difficult to understand why that strategy is optimal and generalize f
by elehack 7y ago
Yes. Bandits will often converge more quickly to the optimal strategy, but it is much more difficult to understand why that strategy is optimal and generalize from the bandit outcomes to predict future performance and performance of other strategies.
It isn't impossible - bandits are seeing adoption in medical trials to avoid precisely the problem discussed - but the standard experiment design and analysis techniques you learn in a decent college statistics class or introductory statistics text no longer apply. That's one of the beauties of A/B testing: while it does require substantial thought to do well, the basic statistics of the setup are very well-understood at this point.
- orasis 7y agoI disagree. I’ve spent a lot of time staring at bandit outcomes and usually they match some sort of intuition of why a variant might be exceptional.
- comicjk 7y agoThat could be post-hoc reasoning, though. It would be interesting to pre-register your hypotheses, or see whether you could tell bandit outcomes from random ones.
- edmundsauto 7y agoIsn't this problem also an issue when people talk about transferring what they learn from one test to another test? That is frequently cited as a benefit of A/B testing.
- orasis 7y agoSure it’s post-hoc reasoning, but it doesn’t matter because I’m not trying to invalidate a hypothesis. I’m looking for variants that win. When I find one that wins I look at it and try to add more of the same flavor to the product. This process works.
- taeric 7y agoThis is literally the logical fallacy. You could get lucky. Maybe you have obvious gains to chase. But bad logical arguments are bad because they never work forever. They are corrupted heuristics that can get you in trouble without critical thinking. Edit: added in forever. Phone dropped some wording I originally had. I think.
- orasis 7y agoCall it a genetic algorithm if you like. I’m looking for incremental wins in a world of infinite possibilities, not truth.
- taeric 7y agoIncremental wins can still lead to dead ends. My phrasing was off in my post. I meant to say that the fallacies aren't that the tactics never work. Just that they can stop working without you really realizing it. A heuristics that can lead you down a dead end. By all means, keep doing it if it is working for you. But don't confuse it as good advice. And stay vigilant.
- orasis 7y agoProducts exist in human reality not some science paper. There are no absolute truths, everything dead-ends eventually. It’s like trying to prove that one set of genes is better than another for future survival - an impossible task.
- taeric 7y agoThis belies a belief that science doesn't reflect the real world. It absolutely does. Again, it may be working in your case. Argument to authority can go a long way. Even ad hom attacks often exist due to a "smell" of the person speaking. It is not, however, logically sound and can easily lead to unsupportable positions. So, take care. And realize that a lot of the damage of poor practices may be tangential. For example, a belief that the real world can not be described by science.
- edmundsauto 7y agoBut for results to generalize or to understand why, the confounders must be accounted for in the randomization. This is really hard to do well -- there are often subtle influences that aren't sufficiently understood how they impact these non-linear systems. What makes someone convert? A million different factors; changing the color of a button in one context doesn't necessarily tell me much about how people would respond to that experience in another context. It's easy to underestimate how complex things are, because we only see some superficial aspects of e.g. a user/software interaction model. This flaw is down to how our brains work -- ref "What you see is all there is".