David Rein


Sorted by New


Everyday Lessons from High-Dimensional Optimization

Which suggests that if you're doing randomish exploration, you should try to shake things up and move in a bunch of dimensions at once rather than just moving along a single identified dimension.


If you can only do randomish exploration this sounds right, but I think this often isn't the right approach (not saying you would disagree with this, just pointing it out). When we change things along basis vectors, we're implicitly taking advantage of the fact that we have a built-in basis for the world (namely, our general world model). This lets us reason about things like causality, constraints, etc. since we already are parsing the world into a useful basis.