How DeepMind's machine learning model MuZero learns to play chess

2019-11-22

DeepMind published a paper in the journal "Science" at the end of last year, detailing AlphaZero, an artificial intelligence system that can teach itself how to master variations of chess, Japanese chess "shogi," and the Chinese game of Go. In competitions of these games, AlphaZero successfully defeated world champions.


However, the prerequisite for AlphaZero's victory is that it must know the rules of the game it is playing. In pursuit of controllable high-performance machine learning models, DeepMind researchers designed MUZERO, a learning model that can predict the most relevant quantities to game decisions, enabling it to achieve industry-leading performance in 57 different Atari games, matching the performance of AlphaZero in Go, chess, and shogi.



Model-Based Reinforcement Learning


Essentially, MuZero receives observation results and converts them into hidden states. The hidden states are iteratively updated through the process of receiving previous states and hypothesized next actions. As explained by DeepMind researchers, MuZero is a form of reinforcement learning, which is the core technology of AlphaZero, where rewards guide the AI agent towards its objectives. This form uses a model that predicts the state transitions of the next step and a reward model that predicts rewards, modeling the given environment as intermediate steps.


Typically, model-based reinforcement learning focuses on modeling the observation stream at the pixel level, but this granularity is computationally expensive in large-scale environments. In fact, there is no existing method to construct a model that facilitates planning in visually complex domains like Atari.



Training and Experiments


The DeepMind team applied MuZero to classic board games Go, chess, and shogi as benchmark challenging planning problems, and used all 57 games in the open-source Atari Learning Environment as benchmarks for visually complex reinforcement learning domains. They trained the system for five hypothetical steps and used 1 million mini-batch training data in board games.


Regarding Go, despite using fewer overall computations, MuZero's performance slightly exceeded that of AlphaZero, indicating a deeper understanding of its position, according to the researchers. As for Atari, MuZero achieved normalized scores at a new level in the mean and median across the 57 games, outperforming the previous state-of-the-art method (R2D2) in 42 out of 57 games and outperforming the best model-based method in all games.


Next, the researchers evaluated a version of MuZero called MuZero Reanalyze, which has been optimized to improve sampling efficiency, and applied it to 75 Atari games with a total of 200 million frames of experience per game. They reported that it achieved a median of 731% in standardized scores, while the previous state-of-the-art model-free methods IMPALA, Rainbow, and LASER achieved 192%, 231%, and 431%, respectively, with significantly reduced training time.


Finally, to better understand the role played by the model in MuZero, the team focused on Go and Pac-Man. They compared the search performance of AlphaZero with a perfect model to the search performance of MuZero using a learned model and found that MuZero can even match the performance of the perfect model when conducting larger searches than the trained search. In fact, it only performs 6 simulations per move, fewer than the number of simulations performed per execution, which is not enough to cover all 8 possible actions in Pac-Man.


The researchers state that MuZero paves the way for learning methods in many practical domains, especially those lacking communication rules or environment dynamics simulators.

RECENT AI NEWS

RECENT AI TOOLS