Multi-game neural network?
-
luigi
- Lives in gote
- Posts: 352
- Joined: Wed Jul 06, 2011 12:01 pm
- Rank: Low
- GD Posts: 0
- Location: Spain
- Has thanked: 181 times
- Been thanked: 41 times
Multi-game neural network?
Out of curiosity: would it be possible to train a neural network (like KataGo) to get good at several games? I'm thinking Go and Hex, for example, or Go and Keil. Would such a net learn to adapt some of the concepts it learned from the first game to the second one?
-
dfan
- Gosei
- Posts: 1599
- Joined: Wed Apr 21, 2010 8:49 am
- Rank: AGA 2k Fox 3d
- GD Posts: 61
- KGS: dfan
- Has thanked: 891 times
- Been thanked: 534 times
- Contact:
Re: Multi-game neural network?
It's certainly possible in theory. You just force your two networks for the two games to share some of their structure. I don't think it would work well in practice, though, for two main reasons:
1) All the AlphaZero-type networks have a structure that is explicitly related to the geometry of the particular game. So for example the AlphaZero network layers are all 19x19. If you wanted to make a common structure you'd have to make it much harder to let the network reason about specific coordinates on the board (or you'd have to keep so much of the game-specific structure that adding the common stuff wouldn't really have an effect).
2) In my opinion, almost all of the common "how to reason about games in general" features reside in the MCTS part of the algorithm, which is already the same from game to game (excepting for the setting of some hyperparameters).
1) All the AlphaZero-type networks have a structure that is explicitly related to the geometry of the particular game. So for example the AlphaZero network layers are all 19x19. If you wanted to make a common structure you'd have to make it much harder to let the network reason about specific coordinates on the board (or you'd have to keep so much of the game-specific structure that adding the common stuff wouldn't really have an effect).
2) In my opinion, almost all of the common "how to reason about games in general" features reside in the MCTS part of the algorithm, which is already the same from game to game (excepting for the setting of some hyperparameters).