|
|
|
|
|
by DieBruderBauer
979 days ago
|
|
The following questions arose to me after watching this poignant video... 1.) Does the AI "know" it's beating the game? 2.) Why can 10 y.o. me know how to progress in the game without taking countless hours? 3.) Wasn't there a Twitch channel that was broadcasting the A.I. beating Pokemon Red? 4.) Is there any correlation between the A.I. used here and the one used by Deep Blue in Chess? |
|
The Deep Blue algorithm just does a search over the tree of possible moves (up to a certain depth), called alpha-beta search, and picks the best position it can find according to a simple evaluation function.
Searching over all these nodes is effective because chess is not random. If it could go over all possible chess positions, it could find the perfect moves to play every time. It doesn't need any training. The problem is that this search grows exponentially as depth increases, and you need to search up to high depths because chess is complex.
For a long time Stockfish, a powerful chess engine, used only a highly optimized alpha-beta search to perform at a very high level.
Now, there is the approach of reinforcement learning for chess like AlphaZero, which is very similar to this reinforcement learning approach.
And there is a mix between the two, like the current Stockfish, which uses a neural network trained on data along with alpha-beta search.