5 ms·

One thing I find interesting to note is that modern day policy gradient deep reinforcement learning methods are incredibly similar to evolutionary algorithms an

One thing I find interesting to note is that modern day policy gradient deep reinforcement learning methods are incredibly similar to evolutionary algorithms and DRL methods have been applied quite successfully to neural architecture search.

tonmoy 9y ago
I wonder why evolutionary algorithms never managed to be that successful.
- uoaei 9y ago
  They're quite successful on complex multiobjective problems, but you need a lot of compute power to cover the solution space sufficiently. Most problems either 1) don't have such stringent requirements on multiobjective optimization (i.e. the problem can be recast with a single objective function, likely one with a gradient), or 2) more typical machine learning algorithms are "good enough" and require less compute power. But for instance designing complex geometries under numerous constraints is well-suited to evolutionary algorithms. EAs also shine when the evaluation for fitness is computationally-intensive, as you can defer the fitness evaluations until absolutely necessary.
  - yters 9y ago
    My understanding is MOEAs are not a whole lot better than random search.