# AlphaGo Zero

AlphaGo Zero is a self-taught Go AI developed by Google DeepMind that mastered the game using only reinforcement learning and self-play, surpassing all previous versions without human data.

AlphaGo Zero is a computer program developed by [Google DeepMind](https://www.wikiprompt.org/wiki/google-deepmind) that plays the board game Go. It is a [neural network](https://www.wikiprompt.org/wiki/neural-network)-based system that learned to play Go entirely from self-play, without any human game data or prior knowledge beyond the game's basic rules. Released in October 2017, it surpassed the performance of its predecessor, AlphaGo, which had defeated top human players, and established a new benchmark for [artificial intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence) in complex strategy games.

Unlike earlier versions of AlphaGo that were initially trained on thousands of human amateur and professional games, AlphaGo Zero started with a blank slate. It used a single neural network that was trained through [reinforcement learning](https://www.wikiprompt.org/wiki/machine-learning), playing millions of games against itself. This approach demonstrated that a system can achieve superhuman performance in a domain without relying on human expertise, a significant milestone in [deep learning](https://www.wikiprompt.org/wiki/deep-learning) research.

## Architecture and Training

AlphaGo Zero's architecture was simpler and more efficient than that of its predecessors. It used a single deep [neural network](https://www.wikiprompt.org/wiki/neural-network) with residual blocks, a design inspired by [ResNet](https://www.wikiprompt.org/wiki/residual-network), to evaluate board positions and select moves. The network was trained using a combination of policy and value outputs, where the policy head predicted the probability of each move and the value head estimated the expected outcome of the game from the current position.

Training proceeded through self-play, where the current version of the network played against itself to generate game data. The network was updated using a variant of the [stochastic gradient descent](https://www.wikiprompt.org/wiki/sgd-variants) optimizer, specifically the Adam optimizer, with a [learning rate schedule](https://www.wikiprompt.org/wiki/learning-rate-schedule) that decreased over time. The system also employed [data augmentation](https://www.wikiprompt.org/wiki/data-augmentation) by rotating and reflecting board positions to increase the diversity of training examples. This process was repeated over millions of games, with the network continuously improving its play.

## Performance and Results

In a series of experiments, AlphaGo Zero achieved remarkable results. After just 72 hours of self-play training, it was able to defeat the previously released AlphaGo Lee, the version that had beaten world champion Lee Sedol in 2016, by a score of 100 games to 0. After approximately 40 days of training, it surpassed AlphaGo Master, a stronger version that had defeated top professionals in early 2017. The final version of AlphaGo Zero, trained for about 40 days, was evaluated to be stronger than all previous AlphaGo versions, with an estimated Elo rating of over 5000, compared to AlphaGo Master's roughly 4800.

These results showed that a system trained without human knowledge could not only match but exceed the performance of systems that had been trained on human games. The self-play approach also proved more computationally efficient, using fewer resources than earlier versions.

## Significance and Impact

The success of AlphaGo Zero had profound implications for the field of [AI](https://www.wikiprompt.org/wiki/artificial-intelligence). It demonstrated that reinforcement learning, combined with deep neural networks, could solve complex problems without the need for human-generated training data. This was particularly notable in a game like Go, which has an enormous search space and was long considered a grand challenge for AI.

The approach used in AlphaGo Zero has influenced subsequent research in [machine learning](https://www.wikiprompt.org/wiki/machine-learning), including applications beyond games. It highlighted the potential of self-play and [curriculum learning](https://www.wikiprompt.org/wiki/curriculum-learning) in training agents, and it contributed to the development of more general reinforcement learning algorithms. The principles have been applied to other domains, such as [generative models](https://www.wikiprompt.org/wiki/generative-ai) and robotics, though with varying degrees of success.

## Legacy and Future Directions

AlphaGo Zero was a key achievement in the history of [DeepMind](https://www.wikiprompt.org/wiki/google-deepmind) and AI research. It was followed by AlphaZero, a more general version that could also play chess and shogi, further demonstrating the versatility of the self-play approach. The ideas from AlphaGo Zero have been incorporated into other AI systems, and they continue to inform research on [neural networks](https://www.wikiprompt.org/wiki/neural-network) and reinforcement learning.

The program also raised questions about the nature of human expertise and the potential for AI to discover novel strategies. In its games, AlphaGo Zero sometimes made moves that were unconventional by human standards, suggesting that AI can find solutions that humans have not considered. This has sparked discussions about the role of AI in scientific discovery and problem-solving.

As of the early 2020s, the techniques pioneered by AlphaGo Zero remain influential, and the system is often cited as a landmark example of what is possible with modern [deep learning](https://www.wikiprompt.org/wiki/deep-learning) and reinforcement learning. Its legacy is seen in the ongoing development of more capable AI systems across various fields.

## See Also

- [Google DeepMind](https://www.wikiprompt.org/wiki/google-deepmind)
- [Neural network](https://www.wikiprompt.org/wiki/neural-network)
- [Machine learning](https://www.wikiprompt.org/wiki/machine-learning)
- [Artificial intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence)

## References

- Silver, D., et al. (2017). "Mastering the game of Go without human knowledge." Nature, 550, 354-359.
- Silver, D., et al. (2018). "A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play." Science, 362, 1140-1144.

---
Source: https://www.wikiprompt.org/wiki/alphago-zero
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-14T04:13:01.678909+00:00