AlphaGo Zero

The network in AlphaGo Zero is a ResNet with two heads. • The stem of the network takes as input a 17x19x19 tensor representation of the Go board. • 8 channels are the positions of the current player's stones from the last eight time steps. (1 if there is a stone, 0 otherwise. If the time step go before the beginning of the game, then 0 in all positions.) • 8 channels are the positions of the other player's stones from the last eight time steps. • 1 channel is all 1 if black is to move, and 0 otherwise. • The body is a ResNet with either 20 or 40 residual blocks and 256 channels. • There are two heads, a policy head and a value head. • Policy head outputs a logit array of size 19 \times 19 + 1, representing the logit of making a move in one of the points, plus the logit of passing. • Value head outputs a number in the range (-1, +1), representing the expected score for the current player. -1 represents current player losing, and +1 winning. ==Training==

AlphaGo Zero's neural network was trained using TensorFlow, with 64 GPU workers and 19 CPU parameter servers. Only four TPUs were used for inference. The neural network initially knew nothing about Go beyond the rules. Unlike earlier versions of AlphaGo, Zero only perceived the board's stones, rather than having some rare human-programmed edge cases to help recognize unusual Go board positions. The AI engaged in reinforcement learning, playing against itself until it could anticipate its own moves and how those moves would affect the game's outcome. In the first three days AlphaGo Zero played 4.9 million games against itself in quick succession. It appeared to develop the skills required to beat top humans within just a few days, whereas the earlier AlphaGo took months of training to achieve the same level. According to Epoch.ai, training cost 3e23 FLOPs. For comparison, the researchers also trained a version of AlphaGo Zero using human games, AlphaGo Master, and found that it learned more quickly, but actually performed more poorly in the long run. DeepMind submitted its initial findings in a paper to Nature in April 2017, which was then published in October 2017. ==Hardware cost==

The hardware cost for a single AlphaGo Zero system in 2017, including the four TPUs, has been quoted as around $25 million. ==Applications==

According to Hassabis, AlphaGo's algorithms are likely to be of the most benefit to domains that require an intelligent search through an enormous space of possibilities, such as protein folding (see AlphaFold) or accurately simulating chemical reactions. AlphaGo's techniques are probably less useful in domains that are difficult to simulate, such as learning how to drive a car. ==Reception==

AlphaGo Zero was widely regarded as a significant advance, even when compared with its groundbreaking predecessor, AlphaGo. Oren Etzioni of the Allen Institute for Artificial Intelligence called AlphaGo Zero "a very impressive technical result" in "both their ability to do it—and their ability to train the system in 40 days, on four TPUs". Mark Pesce of the University of Sydney called AlphaGo Zero "a big technological advance" taking us into "undiscovered territory". Gary Marcus, a psychologist at New York University, has cautioned that for all we know, AlphaGo may contain "implicit knowledge that the programmers have about how to construct machines to play problems like Go" and will need to be tested in other domains before being sure that its base architecture is effective at much more than playing Go. In contrast, DeepMind is "confident that this approach is generalisable to a large number of domains". Chinese Go professional Ke Jie commented on the remarkable accomplishments of the new program: "A pure self-learning AlphaGo is the strongest. Humans seem redundant in front of its self-improvement." ==Comparison with predecessors==

On 5 December 2017, DeepMind team released a preprint on arXiv, introducing AlphaZero, a program using generalized AlphaGo Zero's approach, which achieved within 24 hours a superhuman level of play in chess, shogi, and Go, defeating world-champion programs, Stockfish, Elmo, and 3-day version of AlphaGo Zero in each case. AlphaZero (AZ) is a more generalized variant of the AlphaGo Zero (AGZ) algorithm, and is able to play shogi and chess as well as Go. Differences between AZ and AGZ include: • AZ has hard-coded rules for setting search hyperparameters. • The neural network is now updated continually. • Chess (unlike Go) can end in a tie; therefore AZ can take into account the possibility of a tie game. An open source program, Leela Zero, based on the ideas from the AlphaGo papers is available. It uses a GPU instead of the TPUs recent versions of AlphaGo rely on. == References ==