DeepMind becomes the first AI to learn to play a game by itself

DeepMind is the advanced artificial intelligence (AI) division of Alphabet is known for creating AlphaGo, a program that beat a human professional Go player for the first time.

DeepMind has published a research in the reputed journal Science, about a system that can learn and master any game similar to Chess, Go, or Shogi, without the need for any human intervention.

Why is this a breakthrough?

The biggest factor was that the system, named AlphaZero, learnt to play all three games from scratch without any knowledge of the rules of the game and playing randomly against opponents until it learnt the game.

While machines have been capable of beating humans for a while, case in point, the scintillating wins of Deep Blue over Garry Kasparov. However, Deep Blue was a computer program built specifically to play chess while DeepMind’s achievement phenomenal is that not only it can master any game that is similar to Go, Chess, and Shogi, i.e. games that have all the information available for the system to make a decision. And it can do that by learning how to play against other programs or human, without the engineers having to write a single line of code specific to the gameplay.

To learn each game, AlphaZero’s untrained neural network plays millions of games against ilsef via a process of trial and error calleddeep reinforcement learning, was used to generalize the gameplay engine.

“This ability to learn each game afresh, unconstrained by the norms of human play, results in a distinctive, unorthodox, yet creative and dynamic playing style,”

DeepMind’s AlphaZero team

From completely random play, DeepMind learns from winning, losing and draws to adjust the neural network, understanding the moves that will be advantageous in future.

Crushes other AIs

AlphaZero was pitted against three AI systems to measure it’s effectiveness,

a) Stockfish for Chess
b) Elmo for Shogi
c) AlphaGo Zero for Go

and it absolutely crushed them.

Does this mean AI can now beat humans at any game?

Not yet, while the AI is capable of playing extremely complex games like Chess, etc., it relies on all information necessary to make the next move to be available and visible.

So while AlphaZero can always crush anybody in Chess and games other similar board games, it still is not capable of winning games like Poker which has a dependency on the second player and the second player’s options are hidden.