Artificial Intelligence
Mastering the Game of Go with Deep Neural Networks and Tree Search
Silver et al. · 2016 · Nature
AlphaGo combined deep networks and Monte Carlo tree search to defeat world-class Go players.
Research objective
Build an artificial agent capable of expert-level Go play, long considered out of reach for AI.
Methodology
Combined supervised policy networks trained on human games, a value network for board evaluation, and Monte Carlo tree search. Later self-play refined the policy via reinforcement learning.
Key findings
- Defeated European champion Fan Hui 5-0 and world champion Lee Sedol 4-1.
- Demonstrated that combining learned heuristics with search can outperform purely human strategy.
- Move 37 in game 2 vs. Lee Sedol shocked the Go community as a novel, beautiful move.
Strengths
- Hybrid neural + search approach captured both intuition and deliberation.
- Generalized to AlphaZero (chess, shogi) and AlphaFold (protein folding).
Limitations
- Required massive compute and game-specific engineering.
- Narrow expertise - no transfer to other domains.
Practical implications
- Renewed interest in combining deep learning with explicit search.
- Established DeepMind as a frontier AGI research lab.