Life In 19x19

There are many "strong" AIs already. As a man, I can learn much in fighting against strong AIs. As an AI, however, I can only learn playing Go from self-play games? Why not let a new AI to learn something in fighting against an old AI?

Jardeus wrote:There are many "strong" AIs already. As a man, I can learn much in fighting against strong AIs. As an AI, however, I can only learn playing Go from self-play games? Why not let a new AI to learn something in fighting against an old AI?

Yes, that seems to be a good idea, why not have Leela, Ktago and Elf play each other in traing mode an the learn from those games?

Leela Zero and ELF are no longer training, so they wouldn't benefit. On the other side, KataGo wouldn't benefit because those two bots are enough weaker than KataGo that games with them are unlikely to be useful data for learning.

Leela Zero in the past *did* do something sort of like what you mentioned - it mixed in games from ELF. This worked okay and probably accelerated its run a little, although it's of course not easy to tell for sure what the effect was, to really measure what effect it had, you'd need to run a very large-scale / expensive experiment where you replicate two runs side by side one with the extra data and one without it.

Thank you very much, if I have a bot, I will let it learn from playing against GNUGo, Patchi, Leela, ELF, LeelaZero, KataGo in order. I believe it will be more efficient.