Google LLC’s DeepMind artificial intelligence research unit today detailed MuZero, a deep learning system that can master Go, chess and other games even if it’s not told the playing rules.

MuZero’s impressive game-solving skills stem from its ability to plan what’s the best course of action in a given scenario. It’s this planning capability that’s the main breakthrough from a research standpoint. According to DeepMind, equipping AI models with the capacity to infer what’s the optimal path forward could allow them to make better decisions and become more useful.

“The ability to plan is an important part of human intelligence, allowing us to solve problems and make decisions about the future,” the DeepMind researchers behind MuZero wrote in a blog post today. “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 generalise to new scenarios, a trait we would also like our algorithms to have.”

The issue until now has been that traditional approaches to AI planning come with major limitations. One common approach, known as lookahead search, has been successfully applied to chess but only works if the AI is given detailed information on the environment in which it’s expected to operate. Such information often isn’t readily available in complex real-world situations, a reality that limits the applicability of the technique.

A second, more sophisticated method of implementing AI planning is known as the model-based approach. Researchers teach the neural network to model the environment on its own without being given any pointers by humans. Like the lookahead search method, this approach is difficult to apply to complex situations, though for a different reason. Modeling every detail of even a relatively simple environment, such as the virtual setting of a video game, can be highly challenging to the point of being impractical. 

DeepMind has addressed the issue with its new MuZero system by inventing a third method. It’s a variation of the model-based approach based on the same principals, but rather than mapping out every single detail about the environment, the AI takes into account only the factors strictly relevant to the task at hand. 

“Instead of trying to model the entire environment, MuZero just models aspects that are important to the agent’s decision-making process,” the researchers explained. “After all, knowing an umbrella will keep you dry is more useful to know than modeling the pattern of raindrops in the air.”

DeepMind put its innovation to the test by having MuZero learn to play Go, chess and shogi. The AI was also given the task of completing the Atari57 test, a suite of 57 video games that are commonly used to assess neural network performance. 

“In all cases, MuZero set a new state of the art for reinforcement learning algorithms, outperforming all prior algorithms on the Atari suite and matching the superhuman performance of AlphaZero on Go, chess and shogi,” DeepMind’s researchers wrote. AlphaZero is an earlier AI developed by the unit that defeated the Stockfish chess engine, which had been previously considered the best in the category.

DeepMind is touting the project as a “significant step” toward the AI community’s long-term goal of building general-purpose AI models that can perform a variety of tasks. In the shorter term, MuZero could potentially be applied to other areas besides games. The DeepMind researchers noted that AlphaZero, the earlier AI that beat the Stockfish chess engine, has been repurposed by academics to tackle problems in fields including chemistry and quantum physics. 

Image: DeepMind