3 ms·
> the best brains have overestimated our ability to write useful rules Nah, we (all humans, not just the brightest) can come up with useful rules just fine. Wh
by catnaroek 11y ago
> the best brains have overestimated our ability to write useful rules
Nah, we (all humans, not just the brightest) can come up with useful rules just fine. What we can't do is write useful rules that describe systems we don't understand. This is the crucial part: understanding. Somehow people ended up buying a false promise that computing power can replace human insight as a tool for taming complexity. The art of making models, of coming up with the right simplifying assumptions, of getting the biggest explanatory bang for your complexity buck, has largely been lost.
- AnimalMuppet 11y ago> What we can't do is write useful rules that describe systems we don't understand. Absolutely. But there are other limitations: We have trouble coming up with rule sets that are complete, non-contradictory, and that really span the domain, even on domains that we understand. (You could argue that we only mostly understand those domains...)
- catnaroek 11y agoThese problems aren't insurmountable. > We have trouble coming up with rule sets that are complete, That's okay. Incomplete rules are better than no rules at all. There exist modeling techniques that require making less assumptions about the domain. For example, https://en.wikipedia.org/wiki/Nonparametric_statistics https://en.wikipedia.org/wiki/Nonparametric_statistics can deal with data sets about whose distribution you want to make no assumptions. Also, using an incomplete model and measuring its inadequacy can help us construct a better model. > non-contradictory, Contradictory rules are a fact of life in certain domains. There exist techniques for dealing with them, like https://en.wikipedia.org/wiki/Goal_programming https://en.wikipedia.org/wiki/Goal_programming. > and that really span the domain, You can split the domain into regions that you understand in isolation then patch everything together, possibly using some smoothening transformation to avoid pathological behaviors at the boundaries.
- awinter-py 11y agoEven keeping well-understood models up-to-date with changing reality is a hard problem: both labor-intensive & hard to verify.
- catnaroek 11y agoThis depends on the number of degrees of freedom of your model - the dimension of its parameter space. This is why we have tools like https://en.wikipedia.org/wiki/Tornado_diagram https://en.wikipedia.org/wiki/Tornado_diagram and https://en.wikipedia.org/wiki/Analysis_of_variance https://en.wikipedia.org/wiki/Analysis_of_variance: to keep the model's size under control, while still allowing for adjustments to its parameters in the future.