3 ms·

There's an interesting parallel to be drawn here from prior RL research: "Some evolutionary algorithms keep only the best performers in the population, on th

There's an interesting parallel to be drawn here from prior RL research: "Some evolutionary algorithms keep only the best performers in the population, on the assumption that progress moves endlessly forward. DGMs, however, keep them all, in case an innovation that initially fails actually holds the key to a later breakthrough when further tweaked. It’s a form of “open-ended exploration,” not closing any paths to progress. (DGMs do prioritize higher scorers when selecting progenitors.)" Kenneth Stanley[0], the creator of the NEAT[1]/HyperNEAT (Picbreeder) algorithms wrote an entire book about open-ended exploration, "Why Greatness Cannot Be Planned: The Myth of the Objective". [0]: https://en.wikipedia.org/wiki/Kenneth_Stanley https://en.wikipedia.org/wiki/Kenneth_Stanley [1]: https://en.wikipedia.org/wiki/Neuroevolution_of_augmenting_topologies https://en.wikipedia.org/wiki/Neuroevolution_of_augmenting_t...

paulluuk 1y ago
It's really a choice: do you want to waste compute or do you want to waste potential? While prioritizing higher scorers for selecting progenitors will initially mitigate some of the problems, you will eventually end up with hundreds of thousands of agents that only learned to repeat the letter "a" a million times in a row, which is a huge waste of processing.
tmaly 1y ago
I followed his work on NEAT at the time. It was really cool. But I never imagined we would get to where we are today with AI.
- spwa4 1y ago
  The same could be said of transformers, that only started to perform when scaled up to an absolutely ridiculous degree. I would argue most researchers are of the opinion that any learning system, scaled up enough, would work. I think the limits of machine learning are related to the fact that all ML "knowledge" is secondhand, except talking to humans and already to a much smaller extent programming. Getting AIs to interact with, say, cars, during training is the way forward.

paulluuk 1y ago

It's really a choice: do you want to waste compute or do you want to waste potential? While prioritizing higher scorers for selecting progenitors will initially mitigate some of the problems, you will eventually end up with hundreds of thousands of agents that only learned to repeat the letter "a" a million times in a row, which is a huge waste of processing.

tmaly 1y ago

I followed his work on NEAT at the time. It was really cool. But I never imagined we would get to where we are today with AI.

spwa4 1y ago
The same could be said of transformers, that only started to perform when scaled up to an absolutely ridiculous degree. I would argue most researchers are of the opinion that any learning system, scaled up enough, would work. I think the limits of machine learning are related to the fact that all ML "knowledge" is secondhand, except talking to humans and already to a much smaller extent programming. Getting AIs to interact with, say, cars, during training is the way forward.