Earlier this week, Google’s DeepMind, an Artificial Intelligence (AI) Lab revealed in a new paper that its latest AlphaGo AI software, Zero, managed to teach itself how to play Go, without needing any human help for it. However, the AI still had to be programmed on how to teach itself to perform a specific task.
AlphaGo, the initial version, made headlines after its developers announced that it managed to beat the world’s best Go players. A Chinese board game, Go has an immensely high number of moves, and these are reportedly extremely hard to crack for machines.
AlphaGo AI Zero, the Latest AI Breakthrough
Now, a paper released this Wednesday reveals that there is now an even smarter version of this AI, one that was capable of teaching itself how to play Go. However, the system did not achieve this entirely of its own, in a ‘first step towards taking over the world’.
David Silver, AlphaGo’s lead researchers, explained that this latest model “uses a novel form of reinforcement learning, in which AlphaGo Zero becomes its own teacher.”
He continues by explaining that the neural network base, which initially doesn’t know anything about the game, starts playing Go with itself. It does so thousands of times, while also combining its neural networks with “a powerful search algorithm”.
In doing so, it learns the game and beginning picking the next move, the one that would be best to play. Then, as the current game ends, the AlphaGo AI tries a different neural network, which improves its abilities to predict the best next move, and even the winner of the game.
It continues doing so over and over again, which only helps create a smarter, better AlphaGo AI version, one which is capable of beating its predecessors.
According to the research paper, this latest AlphaGo defeated the AlphaGo that beat the then Go world champion. It did so after just three days of practice, and with a score of 100 to none.
It was also successful in defeating the AlphaGo Master after 40 days of training. However, the AI is still far from being as advanced as its equivalents in science fiction works.
Namely, while this can excel at one task, it wouldn’t really, or at least yet, be capable of more than one at a time, for example.
Nonetheless, many are calling the results presented in the paper published in the journal Nature, a “breakthrough”.
— DeepMind (@DeepMindAI) October 18, 2017
Image Source: Wikimedia