Hey @DeepMindAI, is AlphaGo holonomic or non-holonomic?
What I’m trying to discern is this: if S is the current state of the board, is the next move entirely a function of S, or does the historical path that led to S in that very game a factor?
Oh, yes: these methods treat game playing almost as a perceptual rather than a strategic problem. If you can afford to compute your strategy anew in every turn, you should, unless you are in a team and communication has a nonzero cost.
-
-
What constitutes "strategy"? Go is a perfect information game, so past info shouldn't be needed. On the other hand, AG uses search (MCTS) to plan ahead.
-
I think strategy implies to formulate a plan early on and commit to it, so certain parts of the game tree are being more exhaustively searched than others, based on the commitment instead of the promise they might hold. This reduces decision cost but decreases reward.
- 2 more replies
New conversation -
-
-
It knows that last 8 moves. However, it's almost equivalent to just knowing the current state. There is an advantage of having to recalculate one's strategy at every instance of a board. Unlike humans, there's no strategy to anticipate.
-
So, just to clarify, you are saying that AlphaGo looks at the current board and 8 states back. Interesting. So, in a manner of speaking, it discovers the trajectory of the play.
- 3 more replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.