The AlphaZero algorithm, without a human teacher, mastered chess and the game of shogi in 24 hours so that it beat other champion programs that had previously unconditionally defeated people.
Deep Mind, a company developing algorithms for artificial intelligence (AI), has published data on the results of retraining the AlphaGo algorithm for playing chess and shogi. Previously, AlphaGo was able to beat the world champions in the game of go. Artificial intelligence managed to master new games even faster. The corresponding work is posted on the Cornell University preprint server.
AlphaZero used a reinforcement learning method. It is a subtype of deep machine learning that does not use a human teacher, but exclusively games between two AIs. Although at the beginning both AIs play very weakly, due to their high speed (in comparison with a human teacher) they can play a huge number of games in a short period of time and select well-proven moves and their sequences in certain positions on the board, which gives the ability for algorithms to increase their level extremely quickly.
In this case, AlphaZero achieved a game level higher than any human player in just 24 hours. Then she was allowed to play with Stockfish, the best chess player available, and Elmo, the best shogi player (a highly modified chess set in early medieval Japan).
Despite a very short training period, AlphaZero beat Stockfish 28 times and drew 72 more times. She managed to win against Elmo 90 times, lose 8 times and draw 2 times. These are extraordinarily high results. Both chess and shogi are very different from the go for which AlphaGo was originally created, as well as from each other. For example, "Japanese chess" implies extremely exotic possibilities, such as playing any of the "eaten" pieces of the enemy. The victory of an algorithm created for Go only by playing with a completely identical algorithm of its own means that a human teacher can achieve
IVAN ORTEGA