Today I continued to attend a virtual school on reinforcement
learning, mostly about genetic algorithms. I never thought they still
develop this field and so intensely. Of course, they can't beat deep
learning models in term of accuracy, but sometimes they can do well
enough. For practical applications, that means precisely that. Enough
is enough:) And some things like Pareto-optimisation are even more
natural for evolutionary algorithms. Yes, they usually don't use GPU.
But they are easily parallelisable and thus can be quite fast.
Genetic algorithms are not about the differentiable of continuous
functions, not to mention second-order derivatives. In general, that
sounds amazing, especially if one thinks about an application to
automated provers. When we develop a mathematical theory, it doesn't
matter whether we can produce some particular theorems. What matters
is whether we can consistently move it forward and prove more and
more valuable results. If humans couldn't find proofs of some
statements for hundreds of years, why we think machines should? But
certainly, computers can develop theories in their way, I believe.
That might be helpful.
