1
$\begingroup$

I have a question regarding the ECGA algorithm. This is what I'm currently studying. If ECGA builds a model of the parent pool using the 'minimum description length' (MDL) measure, why cannot we simply choose the model which has the most likely fit of the parents?

Thanks.

$\endgroup$
1
$\begingroup$

Read the answer from a thread on Google, thanks to Jeremy P. Harford. (Posted with permission.)

...You can! Standardized algorithms can always be altered to better fit your needs. They're generalized forms, and specific applications will often have better fits.

But when you start talking about likeliness, be careful. Leveraging Bayes requires recursive optimization regarding which there are (too many) patent trolls and a performance pitfall that can be tricky to avoid. Plus, further empirical sampling can invalidate your algorithm.

And when I say there are too many patent trolls, understand that people who know nothing about Bayesian statistics nor programming are granted patents by other people who know nothing about Bayesian statistics nor programming. It's a mess.

In this case, applying Bayes can create two tiers of algorithms that don't always complete. Depending upon your use case, that can get rather hairy. It's usually a good idea to implement the generic algorithm first, and then alter it as needed -- kind of like when you replace standard containers with use case optimized custom containers.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.