3 ms·
I assume you're being sarcastic; they do point out in the intro and at the end that a deep RL agent could be trained to do the topology selections, but that wou
by gwern 10y ago
I assume you're being sarcastic; they do point out in the intro and at the end that a deep RL agent could be trained to do the topology selections, but that would be more work to get going than some simple evolutionary operators, and is left to future work. Don't worry, I'm sure it'll be A3C all the way down eventually...
- divbit 10y agoWell yes, you could use neural net for the tournament selection, but I was thinking of a much dumber competition that involves a whole lot more distributed GPU power.