>
The preference for the simplest explanation is just a heuristic. ... They care about whether an explanation is right, not simple.
I disagree entirely. Given two explanations that are "right" in that they explain the observations, and one is very simple, and the other involves vast mysterious unknowns, I would say that there is a strong philosophical argument in favour of the simple explanation.
Especially when complex suppositions like a "goal" of evolution have the characteristics that people are drawn to them for emotional reasons, they raise more questions than they answer.
With "right" I meant the explanation which is actually correct. The preference for simple explanations is only a bias for lack of a better criterion to choose; but typically there are lots of more substantive reasons why one explanation is better than the other (e.g., by looking at how it ties in with other theories).
The idea of simplicity is quite arbitrary. In machine learning there is an analogue in Minimum Description Length learning. It turns out that there the kind of representation is crucial; a hypothesis can be minimal under one representation, but not the other; ergo, what is "simple" is not a straightforward question, and therefore the heuristic is somewhat arbitrary. A "goal" to evolution may sound complex to you, but from an information theoretic perspective, the opposite, namely actual, complete randomness, is the most complex thing possible ...
The whole point of this is: we don't really know, the best we can say is "I won't assume things I don't have evidence for." Lastly, your "raise more questions than they answer" is the hallmark of philosophical problems ...
> With "right" I meant the explanation which is actually correct.
Outside of math (and by extension, mathematical models and virtual computer systems), there is no such thing as "correct". We have physical models we tend to rely on because they give good predictions.
But there's no "correct" explanation for physical phenomena. Newton's mechanics were "correct" until Schroedinger and Heisenberg showed they weren't. And I'm sure within a hundred years, we'll have an even better model than modern quantum mechanics.
If you have say two networks that are based on different choices of perspective in coding a MDL with wildly differing lengths and better performance of the larger one, I would be highly suspicious as to whether the the poor performer is properly built from a sensible simplifying metric.
In terms of a general intuition of how the choice of similarly expressive representations affect complexity, one can look to Kolmogorov Complexity. For any two languages used to compute the KC of some string, there is only an additive translative constant for each possible input. The asymptotics of how each language grows its representation with respect to input size by far will dominate such translative overheads. This is sufficient for philosophical purposes. Although in the practical case these choices matter and MDL was created specifically to treat these practicalities.
I disagree entirely. Given two explanations that are "right" in that they explain the observations, and one is very simple, and the other involves vast mysterious unknowns, I would say that there is a strong philosophical argument in favour of the simple explanation.
Especially when complex suppositions like a "goal" of evolution have the characteristics that people are drawn to them for emotional reasons, they raise more questions than they answer.