AlphaGo
By Google DeepMind
AlphaGo is a reinforcement learning-based AI system developed by Google DeepMind that plays the board game Go, best known for defeating world champion Lee Sedol 4 games to 1 in a widely watched 2016 match, a milestone considered years…
Definition
AlphaGo is a reinforcement learning-based AI system developed by Google DeepMind that plays the board game Go, best known for defeating world champion Lee Sedol 4 games to 1 in a widely watched 2016 match, a milestone considered years ahead of prior expert predictions for AI in Go.
Overview
Go had long been viewed as a much harder challenge for AI than chess, due to its enormous search space — vastly more possible board positions than chess — and the difficulty of hand-crafting an evaluation function to judge how good a given board position is. AlphaGo addressed this by combining deep neural networks with Monte Carlo tree search: a "policy network" learned to predict promising moves, and a "value network" learned to estimate the probability of winning from a given position, both trained initially on human expert games and then refined through self-play using reinforcement learning. AlphaGo's public debut came in 2015 when it defeated European champion Fan Hui, followed by its landmark 2016 match against 18-time world champion Lee Sedol in Seoul, which AlphaGo won 4-1. The match drew global attention, both for the technical achievement and for specific moves — notably "Move 37" in game two — that professional Go players described as creative and unconventional in ways that shifted human understanding of the game itself. AlphaGo was succeeded by AlphaGo Zero, which learned to play at superhuman level purely through self-play without any human game data, and then generalized further into AlphaZero, which learned chess, shogi, and Go using the same underlying algorithm. Later work extended this lineage into MuZero, which learned to plan without even being told the rules of the game in advance. AlphaGo is widely regarded as a pivotal moment in AI history, demonstrating that deep reinforcement learning combined with search could master a domain long considered a benchmark of human intuition and strategic depth, and it substantially accelerated mainstream interest in deep learning research.
Key Concepts
- Combines deep neural networks with Monte Carlo tree search
- Uses separate policy and value networks to evaluate moves and positions
- Initially trained on human expert games, then refined via self-play reinforcement learning
- Defeated world champion Lee Sedol 4-1 in a landmark 2016 match
- Succeeded by AlphaGo Zero, which learned without any human game data
- Directly led to the more general AlphaZero and MuZero systems