PAPERS-03-006 · Paper — AlphaGo

The paper

The paper titled “AlphaGo” introduced a new way for computers to play the game of Go, and it quickly became a landmark in artificial intelligence research.

Before this paper

Before AlphaGo, Go was seen as a grand challenge for AI because the game has more possible positions than atoms in the universe. Brute‑force search methods that had succeeded in chess could not handle this enormous complexity. Researchers believed a breakthrough was still many years away, and the field lacked a method to explore the vast search space efficiently.

The key idea

The key insight of AlphaGo was to combine deep neural networks with Monte Carlo tree search. A policy network suggests promising moves, while a value network predicts who is likely to win. This pairing lets the search focus on a few high‑quality lines instead of millions of random possibilities.

How it works

First, the policy network was trained on a large collection of human expert games using supervised learning. Next, the network was refined through reinforcement learning by letting it play against itself and learning from the outcomes. In parallel, a separate value network was trained to estimate the probability of winning from any board position. During actual play, Monte Carlo tree search uses both networks: the policy network narrows the move candidates, and the value network evaluates leaf nodes, allowing the algorithm to look ahead efficiently and choose strong moves.

See it — search guided by learned networks

flowchart LR
  A["Policy net<br/>narrows moves"] --> C["Monte Carlo<br/>tree search"]
  B["Value net<br/>scores positions"] --> C
  C --> D["Choose the<br/>strongest move"]

What they showed

AlphaGo demonstrated its strength by defeating the European champion Fan Hui 5‑0, and later beating the legendary Lee Sedol 4‑1 in March 2016. These results showed that a computer could not only compete with top human players but also dominate them in a game once thought out of reach for machines.

What it changed

The success of AlphaGo changed the AI landscape by proving that pairing learning with search could solve problems previously deemed decades away. It sparked a wave of research into similar approaches, leading to the development of AlphaZero, MuZero, and many modern planning agents. The work showed that deep learning could be integrated with classic search techniques to achieve superhuman performance. Next: AlphaZero.

The one-line takeaway

Combining deep learning with search can turn seemingly intractable problems into solvable ones.

Read it / meet it

Reading the original AlphaGo paper offers a clear view of how the networks and tree search work together, and it’s a great entry point for anyone curious about modern AI. Today, its legacy lives on in newer systems like AlphaZero and MuZero, which you can explore in recent conference proceedings and open‑source implementations.