Does this suggest a problem with using Bayes to generate hypotheses? My impression is that Bayes includes generating hypotheses which look in the most likely places. Are there productive ways of generating accidents, or is paying attention when something weird happens the best we can do?
I’d have sworn that one of the first sequences I read was about improving science by using Bayes to make better choices among hypotheses. On the one hand, my memory is good but hardly perfect, and on the other choosing among hypotheses is related to but not exactly the same as generating hypotheses.
using Bayes to make better choices among hypotheses
That, sure. You can easily argue that Bayesianism provides a better framework for hypothesis testing. But that’s quite different from generating hypotheses.
An example of using Bayes to “generate hypotheses” that’s successful is the mining/oil industry that makes spatial models and computes posterior expected reward for different drilling plans. For general-science type hypotheses you’d ideally want to put a prior on a potentially very complicated space (e.g. the space of all programs that compute the set of interesting combinations of reagents, in your example) and that typically isn’t attempted with modern algorithms. This isn’t to say there isn’t room to make improvements on the state of the art with more mundane approaches.
New family of materials discovered by accident
Does this suggest a problem with using Bayes to generate hypotheses? My impression is that Bayes includes generating hypotheses which look in the most likely places. Are there productive ways of generating accidents, or is paying attention when something weird happens the best we can do?
I think that Bayes is completely silent on how one should generate hypotheses...
http://wiki.lesswrong.com/wiki/Locate_the_hypothesis
I’d have sworn that one of the first sequences I read was about improving science by using Bayes to make better choices among hypotheses. On the one hand, my memory is good but hardly perfect, and on the other choosing among hypotheses is related to but not exactly the same as generating hypotheses.
That, sure. You can easily argue that Bayesianism provides a better framework for hypothesis testing. But that’s quite different from generating hypotheses.
An example of using Bayes to “generate hypotheses” that’s successful is the mining/oil industry that makes spatial models and computes posterior expected reward for different drilling plans. For general-science type hypotheses you’d ideally want to put a prior on a potentially very complicated space (e.g. the space of all programs that compute the set of interesting combinations of reagents, in your example) and that typically isn’t attempted with modern algorithms. This isn’t to say there isn’t room to make improvements on the state of the art with more mundane approaches.