DeepMind is on the forefront of artificial intelligence (A.I.). The computer system it developed, known as AlphaZero, amazed (and terrified) the world in 2017 when it was able to defeat human chess masters at their own game, despite only learning it four hours previous to the matches. That machine has been tested numerous times by even more chess grandmasters, and now people are seeing it do something not yet seen within machines – it is improvising.
Not only was AlphaZero a master at chess, it has also taught itself games such as shogi, commonly called Japanese chess, and Go. In each attempt, AlphaZero was able to beat the previous world champions of the games, who were all human. On DeepMind’s website, developers say they are “thrilled” to see the program developing improvisation and intuition skills, which are not previously known to be in machines.
In a paper published in Science Magazine, it is stated the machine’s ability to master the complicated game of Go and defeat the world champion showed it had use of “deep convolutional neural networks” because it developed a massive knowledge of the game simply by playing it repeatedly to the point the writers of the paper said it has “superhuman performance” in the game.
Computers have been beating humans at chess since 1997, but the addition of shogi, which is far more complicated than chess, and Go, which relies on practice and intuition, shows AlphaZero is able to not only defeat humans at their own games, but ultimately learn how to do it in better and more efficient ways.
When pitted against another chess computer, Stockfish, AlphaZero won 155 of 1,000 matches, with six losses and the rest being draws. Unlike most chess-playing A.I.’s, however, AlphaZero does not prefer to save its pieces, instead opting to sacrifice them for the greater good.
This ability comes from what developers describe as a “neural network with millions of different tunable parameters, each learning its own rules of what is good in chess.” With all of these variables, the machine, much like a human, can look at a situation and know what the best thing to do is.
AlphaZero began with a blank slate mind, developing strategies and tactics based only on the basic rules of the games it plays. It developed its human-like ability to play games based on its experiences.
While many prominent thinkers such as Elon Musk have warned against A.I., citing the possibility such mechanical minds could ultimately lead to human extinction, DeepMind researchers believe studying the way this machine learns how to play games can lead to real issues, such as why proteins become misfolded in Parkinson’s and Alzheimer’s. That protein folding conundrum is ultimately the goal of A.I.’s such as AlphaZero built by DeepMind.