It’s true that other hypotheses better predicting the data would still overcome it in time. But a rigged encoding would still allow for theories which add an epiphenomenal entity to count as less complex; and we definitely don’t want Occam’s Razor to do that. I know there’s no way to guarantee that, say, the universal Turing machine complexity is “fair”, but you can’t dodge the necessity of putting some restrictions on the encoding.
As I said to cousin it, the proof allows for different Razors depending on the specific definition of complexity. This is actually a good thing because we can learn from experience which definition to use in various contexts.
Of course experience doesn’t directly say whether or not epiphenomenal entities exist, but we do discover by experience that the “number of entities” understanding of the Razor often works well, that is, that hypotheses with less entities posited are more probable than ones with more. This would lead us to adopt a Razor that does not count the addition of epiphenomena as less complex.
Would you agree that even if such restrictions are desirable in practice, there is no way to add such restrictions without adding something which is not logically necessary? The original proof is intended to be logically necessary.
As I said to cousin it, the proof allows for different Razors depending on the specific definition of complexity. This is actually a good thing because we can learn from experience which definition to use in various contexts.
At which point, all the difficulty in deciding between theories is transferred to deciding between encodings. Yes we can learn over time which encodings work well in different domains, but at that point, you’re really just putting chunks of theory into your encoding choice.
There is a place for this kind of shorthand: I shouldn’t have to model people in detail to form a hypothesis explaining the presence of a building in a feild. But this is because I know people exist from vast amounts of other data, so I have a justification for the shorthand of ‘a human did it’ in my encoding. It should have a high complexity cost, but it comes packaged with tons of evidence so in most cases it’s practically free.
It’s true that other hypotheses better predicting the data would still overcome it in time. But a rigged encoding would still allow for theories which add an epiphenomenal entity to count as less complex; and we definitely don’t want Occam’s Razor to do that. I know there’s no way to guarantee that, say, the universal Turing machine complexity is “fair”, but you can’t dodge the necessity of putting some restrictions on the encoding.
As I said to cousin it, the proof allows for different Razors depending on the specific definition of complexity. This is actually a good thing because we can learn from experience which definition to use in various contexts.
Of course experience doesn’t directly say whether or not epiphenomenal entities exist, but we do discover by experience that the “number of entities” understanding of the Razor often works well, that is, that hypotheses with less entities posited are more probable than ones with more. This would lead us to adopt a Razor that does not count the addition of epiphenomena as less complex.
Would you agree that even if such restrictions are desirable in practice, there is no way to add such restrictions without adding something which is not logically necessary? The original proof is intended to be logically necessary.
At which point, all the difficulty in deciding between theories is transferred to deciding between encodings. Yes we can learn over time which encodings work well in different domains, but at that point, you’re really just putting chunks of theory into your encoding choice.
There is a place for this kind of shorthand: I shouldn’t have to model people in detail to form a hypothesis explaining the presence of a building in a feild. But this is because I know people exist from vast amounts of other data, so I have a justification for the shorthand of ‘a human did it’ in my encoding. It should have a high complexity cost, but it comes packaged with tons of evidence so in most cases it’s practically free.