Contents
Is there a neural network for AlphaZero chess?
Initially AlphaZero was something of a mystery to me. Like everyone else, I knew it made use of a neural network, but to me that didn’t mean much. Add in the fact that it learned from selfplay and you might as well have called it magic. But since it was humans that designed AlphaZero, we shouldn’t despair of trying to understand it.
What happens to a neuron in a chess network?
One thing to notice is that what a neuron outputs potentially depends on every other neuron in the network before it, which allows the network to capture subtleties, like in chess where White’s castled king is safe, but after h3 the assessment changes as Black can open the g-file with g7-g5-g4.
How is a chess NN might be constructed?
How a chess NN might be constructed. The “Input Layer”, is, of course, where the position is inputted into the network. Since we’re using bitboards you can put them in as 64×12 stacks (8×8 squares, times 12 pieces) or as a long string of 1s and 0s (12*64=764 inputs) The second way is how I do it here.
Can a chess network be used in a chess engine?
This just means that a neural network is given a chess position, and is designed to output a move and an evaluation. Math-folk might recognize that what we require is a function with a domain of chess positions, and a range of legal moves and evaluations. Obviously, a physical chessboard can’t be used in an engine.
What kind of dataset does AlphaZero use?
Every single one of them (including the original AlphaGo) uses ridiculously large precomputed tablebases of moves, professional datasets of “well-played games”, and carefully crafted heuristic functions with tons of hacky edge-cases.
Why is so much memory needed for deep neural networks?
RNNs are a special type of DNN that allow cycles in their structure to encode behaviour over sequences of inputs. For RNNs, re-computation has been shown to reduce memory by a factor of 20 for sequences of length 1000 with only a 30% performance overhead.
Why are there so many moves in AlphaZero?
There are about 150-250 moves on average playable from a given game state. The reason for the slow progress of DFS is that when estimating the value of a given state in the search, both players must play optimally, choosing the move that gives them the best value, requiring complex recursion.