DeepMind’s MuZero AI masters games without knowing the rules – The Burn-In

Its no secret that artificial intelligence (AI) programs are good at mastering games. From chess to Go to StarCraft II, humans are often no match for the computer. One thing that previous AI game engines have in common is that they were taught the rules of each game before learning to play it.

Now, researchers from DeepMind have taken things a step further, Engadget reports. The Alphabet subsidiary just announced that its MuZero AI is capable of playing a variety of games (and crushing human players) without being told the rules. Thats an incredibly impressive achievement and demonstrates how far the field of AI has come in the past few years.

When DeepMind unveiled its original AlphaGo AI in 2016, the world took notice. The crafty algorithm was capable of beating the worlds best human players in the ancient strategy game, Go.

It then topped the original version by creating AlphaGo Zero, an AI capable of learning to play Go simply by playing against itself. That was a step up from the original version, which learned by watching people play.

From there, DeepMind branched out. It taught its AlphaZero algorithm to play Go, chess, and shogi. Being able to master all three games was another impressive accomplishment.

The companys latest algorithm, MuZero, is a sort of combination of every iteration. Not only did it forgo the need to hear the rules of the three games noted above, it also knows how to play a suite of classic Atari titles. The algorithm learned to play all of the games on its own and is even better at them than any other version that DeepMind has created so far.

This addresses a problem that researchers have been pondering for some time, notes Engadgets Igor Bonifacic.

Its difficult to train an AI to strategize and plan ahead when it doesnt know the rules of the game thats being played. In the real world, there is no such thing as a simple set of rules. Thats what makes this breakthrough so important for the world of AI.

DeepMinds MuZero combines aspects of two prediction models that researchers have used in the past to try and solve the problem of predicting something that isnt governed by a strict set of rules.

This means that MuZero is thinking a lot more like a human than a computer. It must rule out factors that arent likely to impact the overall outcome to focus on the ones that will.

DeepMind likens this to looking at a sky full of storm clouds. In a blog post, the company says, For example, if we see dark clouds forming, we might predict it will rain and decide to take an umbrella with us before we venture out. Humans learn this ability quickly and can generalize to new scenarios, a trait we would also like our algorithms to have.

MuZero represents a big leap forward in AI technology. DeepMind notes that it has potential applications in fields like robotics, industrial systems, and other messy real-world environments.

Continue reading here:

DeepMind's MuZero AI masters games without knowing the rules - The Burn-In