AI Taught Play Go Beat AI Champion?

Yes, an AI trained to play Go has beaten an AI champion. In 2017, AlphaGo Master, a program developed by DeepMind, defeated the world’s top-ranked human player, Ke Jie, in a three-game match. This article explains how AI systems learn to play Go, how they compete against each other, and what it means for the future of AI.

How AI Learns to Play Go

Go is an ancient board game with a 19×19 grid. The number of possible board positions is more than the number of atoms in the observable universe, so traditional AI methods that search every move fail. Instead, modern Go AI uses deep neural networks and reinforcement learning.

The neural network evaluates board positions and suggests moves. It is trained in two phases:

• Supervised learning: The network learns from millions of moves played by human experts. It predicts the next move a human would make.
• Reinforcement learning: The AI plays against itself millions of times. It receives a reward for winning and adjusts its network to favor moves that lead to victory.

This self-play loop is what allows the AI to surpass human knowledge. It discovers novel strategies that humans never used, sometimes called “creative” moves.

AlphaGo vs. AlphaGo: The Self-Play Revolution

After AlphaGo beat the European champion Fan Hui in 2015, its successor AlphaGo Zero started from scratch. It had no human game records. It only knew the rules of Go. By playing against itself, it reached a level that outperformed the original AlphaGo within days.

When two AI versions of AlphaGo play each other, the result is not predetermined. The stronger version wins more often, but the weaker one can occasionally win if it plays a surprising move. This is similar to human chess champions: the higher-rated player wins most games but not all.

In 2017, AlphaGo Master (a refined version) played 60 online games against top human professionals and won all of them. It then faced Ke Jie in a formal match and won 3-0. After that match, DeepMind retired AlphaGo from public competition.

What Makes One Go AI Better Than Another?

Several factors determine which AI wins in a head-to-head match:

• Network architecture: More layers and parameters can capture more complex patterns, but require more computation.
• Training time: More self-play games usually lead to stronger play, but with diminishing returns.
• Search algorithm: Monte Carlo Tree Search (MCTS) is used to evaluate the best move by simulating many future games. The number of simulations per move affects strength.
• Hardware: Faster processors and more memory allow deeper search and faster training.

In practice, the strongest Go AI today (as of 2024) is likely a successor to AlphaGo, such as KataGo or Leela Zero, which are open-source projects. These programs have beaten each other in various tournaments, and the winner often depends on the specific version and hardware.

How to Watch an AI vs. AI Go Match

If you want to see two AIs play, you can:

1. Download an open-source Go engine like KataGo or Leela Zero.
2. Install a graphical interface like Lizzie or Sabaki.
3. Load two engines and set them to play against each other.
4. Adjust the number of playouts (simulations per move) to control strength.

You can also watch online platforms like KGS or OGS, where bots regularly play games. These matches are often streamed on YouTube with commentary.

Common Misconceptions About AI vs. AI

One common myth is that if an AI plays itself, it will always end in a draw because both sides are identical. That is false. Even identical programs can produce different moves because of random elements in the search (e.g., random tie-breaking). Over many games, the stronger version (or the one with more playouts) will win more often.

Another misconception is that AI vs. AI games are boring or meaningless. In fact, they are used to test new algorithms and to discover new Go strategies. For example, AlphaGo’s famous “Move 37” against Lee Sedol was a move that humans initially thought was a mistake but later recognized as brilliant. Such insights come from AI self-play.

Frequently Asked Questions

Can a human beat a top Go AI?

As of 2024, no human can consistently beat the strongest Go AI. The gap is too large. Top professionals occasionally win a single game if the AI makes an unusual mistake, but that is rare. The AI’s ability to evaluate positions and search deeply gives it a decisive edge.

How long does it take to train a Go AI?

Training time varies. AlphaGo Zero used 40 days on specialized hardware (TPUs) to reach superhuman level. Open-source projects like Leela Zero have trained using distributed computing from volunteers, taking months. On a single consumer GPU, training a strong Go AI from scratch could take years, but you can download pre-trained networks.

Is Go AI stronger than chess AI?

Both are superhuman, but they are not directly comparable. Go has a much larger branching factor, so it was considered harder. Modern Go AI uses similar techniques to chess AI (neural networks + search), but Go requires more computation. Both have surpassed the best humans by a wide margin.

What happens if two identical AIs play each other?

If they are truly identical (same network, same search, same hardware, no randomness), the game would be deterministic and always end the same way. But in practice, there is always some randomness (e.g., in the search), so games vary. If one AI has more playouts or a slightly different version, it will usually win.

Can I train my own Go AI?

Yes, but it requires significant computational resources and programming knowledge. You can use open-source frameworks like TensorFlow or PyTorch to implement a simple Go AI. For a strong bot, you would need to train for weeks on a powerful GPU. Alternatively, you can use pre-trained models and just run them.

Limits of AI vs. AI Comparisons

When comparing two Go AIs, the result is only valid under the specific conditions of the match: hardware, time per move, and version. A bot that wins on one machine might lose on another if the opponent has more computing power. Also, the outcome does not tell you which AI is “smarter” in general; it only tells you which played better on that day.

Furthermore, AI vs. AI matches do not necessarily reflect how the AI would perform against a human. Some AIs are tuned to exploit weaknesses in other AIs, which might not apply to humans. So the result of an AI vs. AI match is interesting but not the final word on overall strength.

Related

• Survival Horror with Fast Shooter Combat: Top Picks
• Power User Guide htop: Master Linux Process Management
• Fanless Industrial PCs vs Traditional PCs for Factory Floors

in t | Time (minutes) | min | 30 | 0 | 600
in d | Distance (km) | km | 10 | 0 | 100
out pace | Average pace (min/km) | min/km | t / d | 2
out speed | Average speed (km/h) | km/h | d / (t / 60) | 2