Computer Go
Computer Go is the project of making a machine play Go well.
6 frameworks, from 1965 to today · Go · Games
Overview
Computer Go is the project of making a machine play Go well1. It matters out of proportion to the game's modest commercial footprint because Go resisted computation longer than almost any other classic board game. Chess fell to Deep Blue in 1997; Go's best programs were still losing to strong amateurs nearly two decades later. The reason is structural. Go's branching factor — the number of legal moves at a typical position — is around 250, against roughly 35 in chess, and games run to 200–300 moves. Brute-force search, the engine of chess computing, drowns in that1. Worse, Go positions are hard to evaluate: unlike chess material, the value of a Go position is a subtle judgment about territory, influence, and life and death that strong players read intuitively12. Computer Go's history is the history of trying to solve those two problems — search and evaluation — and of the moment when machine learning dissolved the distinction between them.
Early programs and the heuristic era (1960s–1980s)
The first Go programs appeared in the 1960s, written by researchers testing ideas about machine intelligence1. Early efforts played on small boards and were weak, but they established the basic architecture that would dominate for decades: a program that generated candidate moves, evaluated the resulting positions with hand-coded rules, and searched a few moves ahead. The evaluation was the hard part. Programmers encoded Go knowledge as heuristics — rules of thumb such as "extend from a wall," "capture a stone in atari," "prefer moves near the edge in the endgame."1 This is what the field calls knowledge-based Go: the attempt to write down, in code, what human players know1.
The approach had a natural appeal and a natural ceiling. Go knowledge is vast, much of it tacit, and often context-dependent; a heuristic that helps in one position misleads in another. Programs improved through the 1970s and 1980s, and the first computer Go tournaments began — the US Computer Go Championship dates from the late 1970s, and international events followed. But progress was slow, and the gap to human strength remained enormous.
Alpha-beta search and its limits (1980s–1990s)
The standard search method of the era was alpha-beta search, a refinement of minimax that prunes branches of the game tree that cannot affect the result3. In chess it was transformative. In Go it struggled1. Alpha-beta search needs a reliable evaluation function to prune well, and Go evaluation was unreliable; it also needs a manageable branching factor, and Go's was not1. Programs that searched deeply in Go often searched deeply into nonsense. The best programs of the 1990s were still far below dan-level amateur strength, and many researchers concluded that Go would not yield to the methods that had conquered chess1.
A partial exception was the endgame. In the late game, when the board is nearly settled, the position becomes tractable, and programs using combinatorial game theory and alpha-beta methods could play the yose (endgame) at respectable strength. This suggested a pattern that would recur: Go is not uniformly hard, and progress came from attacking its tractable parts.
Monte Carlo methods and the MCTS revolution (2006–2010s)
The turning point came from an unexpected direction. Instead of evaluating positions with hand-coded rules, a program could sample them: play out the position many times with random or lightly guided moves to the end of the game, and average the results1. This is Monte Carlo evaluation, and its great virtue is that it needs no evaluation function at all — the playouts supply the estimate1. Its weakness is variance: random playouts are noisy, and early Monte Carlo Go programs were weak1.
The decisive idea was to combine sampling with search. Monte Carlo tree search (MCTS) builds a search tree incrementally, allocating more playouts to promising branches and balancing exploration against exploitation14. The key algorithm, Upper Confidence Bounds applied to Trees (UCT), was introduced in 2006 by Levente Kocsis and Csaba Szepesvári, and applied to Go almost immediately by Rémi Coulom and others. The effect was dramatic. Within a few years, MCTS programs leapt past the previous generation; by around 2008–2010, programs like MoGo and Crazy Stone were playing at strong amateur level on 9×9 boards and improving rapidly on 19×191. MCTS did not need to know what a good Go move was; it needed only to try many moves and let the statistics sort them out.
MCTS programs still benefited from Go knowledge — better playout policies, patterns learned from game records, and hand-tuned move priors all helped — but the balance had shifted1. The search was doing the work that heuristics had done before.
Neural networks and the AlphaGo breakthrough (2015–2017)
MCTS had a remaining weakness: the playouts were still guided by relatively crude policies, and the evaluation of a position was still a statistical average over noisy simulations1. The next step replaced both with learned functions. Neural-network evaluation and policy learning means training a network to do two jobs: a policy network that predicts which moves a strong player would consider, and a value network that estimates who is winning from a position2. The policy network narrows the search; the value network replaces the noisy playout average with a learned judgment2.
The landmark was AlphaGo, developed by DeepMind and announced in 2015. Its first version combined MCTS with policy and value networks trained on human game records, and in October 2015 it defeated Fan Hui, the European professional champion — the first time a computer beat a professional Go player without handicap12. In March 2016, AlphaGo defeated Lee Sedol, one of the strongest players in the world, four games to one. The match was a global event, and Lee's single win in game four — a brilliant, unexpected move — became one of the most discussed games in Go history.
AlphaGo's successor, AlphaGo Zero (2017), removed the human game records entirely15. It learned from self-play reinforcement learning: starting from random play, it played against itself, and the results of those games were used to improve the policy and value networks, which in turn generated better self-play. This is the idea that a system can bootstrap its own skill without human examples, using only the rules of the game and the outcome of wins and losses65. AlphaGo Zero surpassed the version trained on human data, and its successor AlphaZero extended the method to chess and shogi15. The lesson was that in Go, human knowledge, however deep, was not the ceiling — it was a starting point that could be surpassed.
The engine era and the present
Since 2017, the strongest Go players in the world have been programs. Open-source engines such as Leela Zero, built on the AlphaGo Zero approach and trained by distributed volunteer computing, and KataGo, which added efficient search and support for variable rules and komi, became widely available. Professional players now study engine games and engine evaluations; the joseki (corner sequences) and opening patterns of the pre-engine era have been revised, and some have been abandoned. The engine's judgment — its win-rate estimates and its preferred moves — has become a reference point in professional study, much as chess engines became for chess.
The field's open questions are no longer about whether machines can play Go well. They concern efficiency (how little computation is needed to reach strong play), generalization (how the methods transfer to other games and to real-world problems), interpretability (what the networks have learned and why they choose the moves they do), and the human side: how professional Go culture absorbs and adapts to a game whose strongest authorities are now machines. The history of Computer Go is, in that sense, still being played out.
References
- Computer Go (introduction, Minimax tree search, Challenges for strategy and performance for classic AIs, Early decades, Knowledge-based systems, System design, Overview and history, Monte-Carlo methods, 2007–2014: Monte Carlo tree search, 2015 onwards: The deep learning era), Wikipedia
- AlphaGo (History, introduction), Wikipedia
- Alpha–beta pruning, Wikipedia
- Monte Carlo tree search, Wikipedia
- AlphaGo Zero, Wikipedia
- Self-play, Wikipedia
The frameworks of Computer Go
Each framework is a way of working the field has used: when it held, and the question it answers.
- 1965-2005
- 1980-2000
- 2006-2017
Monte Carlo tree search
How can random sampling guide search in complex games?
- 2015-Present
Neural-network policy and value learning
How can neural networks learn to evaluate and select moves?
- 2017-Present
Open-source engine ecosystem
How can strong Go engines be built and shared openly?
- 2017-Present
Self-play reinforcement learning
Can a system learn complex skills without human examples?
How the ideas connect
Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.
Monte Carlo tree search replaced Alpha–beta pruning
Monte Carlo tree search replaced alpha-beta search as the dominant search paradigm in Computer Go. Alpha-beta's reliance on reliable evaluation and manageable branching made it ineffective for Go, and MCTS overcame both limitations through sampling.
Carried by Rémi Coulom.
Neural-network policy and value learning influenced by Monte Carlo tree search
AlphaGo combined Monte Carlo tree search with neural-network policy and value networks, using MCTS as the search engine guided by learned evaluations. The MCTS framework provided the integration mechanism for the neural networks.
Carried by DeepMind; AlphaGo (2015).
Self-play reinforcement learning grew out of Neural-network policy and value learning
AlphaGo Zero extended the neural-network approach by removing human game records and training solely through self-play reinforcement learning. It grew directly out of the policy and value network architecture of AlphaGo.
Carried by DeepMind; AlphaGo Zero (2017).
Open-source engine ecosystem influenced by Self-play reinforcement learning
Open-source engines like Leela Zero and KataGo implemented the self-play reinforcement learning methods of AlphaGo Zero, using distributed volunteer computing to train strong Go programs. The self-play paradigm became the foundation of the open-source engine ecosystem.
Carried by Leela Zero; KataGo.
Alpha–beta pruning influenced by Knowledge-based Go
Knowledge-based Go programs supplied the hand-coded evaluation functions that alpha-beta search required for pruning. Researchers applied alpha-beta to Go using these heuristic evaluations, though the unreliability of the evaluations limited its success.
Monte Carlo tree search influenced by Knowledge-based Go
MCTS programs incorporated Go knowledge from the heuristic era, using hand-tuned move priors and pattern-based playout policies to guide simulations. This knowledge improved the efficiency of Monte Carlo tree search without being its foundation.
Beyond the field
Where Computer Go's ideas came from in other fields, and what they led to there: the strongest 12 of 17 links.
Neural-network policy and value learning influenced Neural-network opening assessment in Opening Theory
Neural-network policy and value learning developed in computer Go, culminating in AlphaGo and AlphaZero, was transferred to chess and produced the neural-network opening assessments that replaced handcrafted evaluation.
Carried by AlphaGo; AlphaZero.
Self-play reinforcement learning influenced AlphaZero in Computer Chess
AlphaZero is the demonstration that pure self-play reinforcement learning with a neural network can master chess, Go and shogi, extending the approach to the more general self-play framework.
Self-play reinforcement learning influenced by Self-play in Game Playing Ai
AlphaZero's self-play reinforcement learning built directly on the long tradition of self-play training in game-playing AI, training a network solely by playing against itself from random initial play.
Knowledge-based Go influenced by Symbolic artificial intelligence in Philosophy Of Artificial Intelligence
Knowledge-based Go was built within the symbolic AI paradigm, encoding expert Go knowledge as hand-crafted rules, patterns and heuristics in the manner of other symbolic expert systems.
Knowledge-based Go influenced by Knowledge-Based Chess AI in Game Playing Ai
Early computer Go programs adopted the knowledge-based, hand-crafted evaluation and pattern approach that had been developed in computer chess, with researchers drawing directly on chess AI techniques and sometimes working across both games.
Knowledge-based Go influenced Database-Empirical Joseki Study in Joseki Theory
Knowledge-based Go research supplied pattern databases and joseki libraries that later empirical joseki study drew on when mining game records for opening statistics.
Alpha–beta pruning influenced by Search-and-evaluation engine in Computer Chess
Search-and-evaluation chess engines used alpha-beta pruning as their core tree-search method, combining it with static evaluation at the leaves.
Alpha–beta pruning replaced by AlphaGo in Life And Death
AlphaGo defined itself against alpha-beta search, which had failed to reach strong play in Go; it replaced deep alpha-beta evaluation with Monte Carlo tree search guided by neural networks. The 2016 Nature paper presents this as the alternative to prior search-based Go programs.
Carried by AlphaGo (2016).
Alpha–beta pruning replaced by Monte Carlo Tree Search in Life And Death
MCTS in Computer Go arose as a response to the failure of alpha-beta search with hand-tuned evaluation in Go, where branching is too wide for deep pruning. The 2006 rollout-based programs abandoned alpha-beta in favour of statistical simulation.
Carried by Rémi Coulom.
Neural-network policy and value learning influenced by Neural-Network Engine Analysis in Backgammon
TD-Gammon showed that a neural network could be trained by self-play with value estimates to play backgammon at strong level, providing the template for neural value and policy learning in games that AlphaGo's makers built on.
Carried by Gerald Tesauro.
Neural-network policy and value learning influenced by Policy gradient method in Reinforcement Learning
AlphaGo's policy network was trained by policy-gradient methods of reinforcement learning, optimising move probabilities directly against game outcomes.
Neural-network policy and value learning influenced by Value-Function Methods in Reinforcement Learning
The value network in AlphaGo was trained with value-function methods from reinforcement learning, estimating state values to guide search.

Computer Go is a planet in Nousfarer
Its city has a tower for each of these 6 frameworks. Land, walk in, and climb one concept at a time.
Land on Computer GoRelated fields
Fields its ideas are linked to
- Computer Chess 3 links
- Game Playing Ai 3 links
- Reinforcement Learning 2 links
- Life And Death 2 links
- Knowledge Representation Reasoning 1 link
- Planning And Search 1 link
- Opening Theory 1 link
- Go 1 link
- Joseki Theory 1 link
- Philosophy Of Artificial Intelligence 1 link
- Backgammon 1 link