Can Alpha Zero General handle games with imperfect information? #310
-
|
For example, games like poker where some information is unknown, or games where the next card to be drawn from a deck is unknown. A neural network shouldn't (generally) care if some information is unknown, but MCTS implementations may depend on determinism, especially if they need to run on the GPU. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
The Alpha Zero algorithm (not only this implementation) is meant to work on perfect games. There are more recent algorithms like ReBeL which are designed to support imperfect games such as poker. However, by "torturing" a bit the Alpha Zero algorithm, I find it possible to support slightly random games (most information is predictable, only a small proportion is not - so not poker) and where same information is shared by all players. Btw it is also possible to make it support more than 2 players. But that is still kind of a hack. |
Beta Was this translation helpful? Give feedback.
The Alpha Zero algorithm (not only this implementation) is meant to work on perfect games. There are more recent algorithms like ReBeL which are designed to support imperfect games such as poker.
However, by "torturing" a bit the Alpha Zero algorithm, I find it possible to support slightly random games (most information is predictable, only a small proportion is not - so not poker) and where same information is shared by all players. Btw it is also possible to make it support more than 2 players. But that is still kind of a hack.