Life And Death
In Go, most stones placed on the board will eventually belong to a group that either lives or dies.
9 frameworks, from 1600 to today · Go · Games
Overview
In Go, most stones placed on the board will eventually belong to a group that either lives or dies. Life and death is the branch of Go knowledge concerned with that question: whether a group of stones can be kept on the board, whether it can be killed, and how many moves each side must spend to settle the matter. A group with two separate enclosed spaces—eyes—cannot be captured, because the opponent can never fill the last liberty inside either one1. A group with one eye, or with a single large enclosed space that the opponent can reduce to one eye, is dead if the surrounding stones are strong enough. The whole subject turns on this asymmetry: the defender needs two eyes, the attacker needs only to prevent the second.
What makes life and death a distinct field rather than a corner of tactics is that it is where reading, shape intuition, and counting meet. A local position may be settled by a forced sequence of a dozen moves, or it may be unsettled—dependent on who moves first, on ko threats elsewhere, or on the exact number of liberties in a capturing race. Practitioners have spent centuries building vocabularies, problem collections, and theories to handle these cases, and in the last two decades computers have transformed what players believe about them.
Early Problem Collections and the Birth of a Discipline
Life-and-death problems are among the oldest recorded Go materials. Chinese and Japanese manuscripts from the medieval and early modern periods include diagrammed positions asking the reader to find the killing or saving move. The Japanese term tsumego (詰碁, literally "packed Go") came to denote these problems, and the genre became a central teaching tool of the four hereditary Go houses—Honinbo, Inoue, Yasui, and Hayashi—that dominated Japanese professional Go from the seventeenth century.
The houses trained their students largely through problem study. A young player worked through collections of increasing difficulty, learning standard killing sequences by repetition before understanding them analytically. This was not mere memorization: the collections encoded the accumulated results of generations of analysis, and a player who had internalized them could recognize a life-and-death situation at a glance and know which moves were worth reading. The most celebrated classical collection, the Gengen Gokyo (玄玄碁経), compiled in China and transmitted to Japan, contains problems attributed to earlier centuries and remained a reference for professionals into the modern era.
The problem collection as a form persisted because it solved a real pedagogical problem: life-and-death reading is expensive to learn from games alone, since a single game may contain only a few genuinely difficult local positions. A curated problem isolates the skill and lets the student practice it hundreds of times.
Classical Shape Knowledge and the Vocabulary of Groups
Alongside the problems grew a body of shape knowledge—recognized configurations of stones with known life-and-death status2. Terms like nakade (the vital point inside an enclosed space), hane (a diagonal move around an opponent's stone), and the distinction between six-space and seven-space eye shapes became standard professional vocabulary. The key classical result is the nakade principle: in a confined space, the attacker plays at the vital point to reduce the defender's eye space, and the number of spaces determines whether the defender can make two eyes1.
This knowledge was codified in the Japanese house tradition and passed down through teaching games and problem sets. Its practical value was enormous: a player who knew that a certain enclosed shape was dead as it stands could save the reading time and spend it elsewhere, and a player who knew that a shape was alive could ignore an opponent's probing move. Shape knowledge functioned as a compression of analysis—a way of storing the results of deep reading so they could be recalled instantly.
Seki: The Theory of Mutual Life
Not every group resolves into life or death1. A seki (セキ, "barrier") is a position in which two groups, typically belonging to different players, share liberties in such a way that neither can capture the other without self-capture13. The classical example is two groups each with one eye, adjacent to each other, sharing a liberty that neither can fill. Seki theory—the study of when such mutual-life positions arise and how they affect the surrounding position—became a recognized part of life-and-death analysis because seki positions are often unsettled in a way that ordinary life and death is not: the status depends on the liberties of both groups and on the order of play elsewhere.
Seki matters because it changes the counting of the game3. A group that lives in seki occupies points that are neither territory nor prisoners, and the surrounding stones may be stronger or weaker than they appear. Classical texts treated seki as a special case requiring separate analysis, and the vocabulary of seki and mutual life entered the professional lexicon.
Shape Theory (Katachi) and the Japanese Analytical School
In the twentieth century, Japanese professionals began to systematize shape knowledge into a theory of katachi (形, "form" or "shape"). The idea was that good shape—efficient, flexible, hard to attack—could be recognized and taught as a set of principles rather than a list of memorized positions2. The most influential figure was Fujisawa Shuko (1925–2009), whose writings on shape and on the vital point of a position shaped a generation of players. Fujisawa's approach emphasized that life-and-death status is often determined by the shape of the surrounding stones, not just the enclosed space: a group with a large eye space can still die if its outside shape is weak, and a group with a small eye space can live if its outside shape is strong.
This was a genuine advance over pure problem memorization. It gave players a way to evaluate a position they had never seen, by asking whether the shape was efficient and whether the vital point was accessible. The Japanese analytical school, including writers like Ishida Yoshio and the problem composer Kano Yoshinori, produced systematic treatments of life and death that organized positions by shape type and by the number of eye spaces.
Deep Reading (Yomi) and the Brute-Force Tradition
The classical method for settling a difficult life-and-death position was yomi (読み), deep reading: the player mentally plays out every plausible sequence, checking liberties and captures, until reaching a position where the status is clear3. Reading is the core skill of professional Go, and life-and-death problems are its training ground.
The brute-force tradition—reading every branch to a conclusion—was the dominant professional method through the twentieth century. Its limitation is combinatorial: a position with many liberties and many possible moves can generate a tree too large to read exhaustively in the time available. Professionals compensated with selective reading: using shape knowledge to prune the tree, reading only moves that are consistent with good shape and with the vital-point principle. The best readers, such as Sakata Eio (1920–2010), were famous for the depth and accuracy of their reading in complex capturing races.
The Computer Revolution: Monte Carlo Tree Search and Neural Evaluation
For decades, computer Go programs were weak at life and death4. The branching factor of a local position is large, and the evaluation of a position—whether a group is alive or dead—is hard to encode as a simple heuristic5. Early programs used hand-crafted rules and shallow search, and they were easily beaten by strong amateurs in tactical fights.
The turning point came with Monte Carlo tree search (MCTS), a method that evaluates positions by playing out many random games from them and averaging the results. MCTS was applied to Go in the mid-2000s and produced a dramatic improvement in playing strength, particularly in tactical positions. The key insight was that random playouts, while individually weak, aggregate into a useful estimate of the position's value, and the tree search focuses computation on the most promising moves.
The second turning point was the application of deep neural networks to Go, culminating in AlphaGo's defeat of Lee Sedol in 2016456. AlphaGo combined a policy network (which suggests moves) with a value network (which estimates the position's outcome), trained first on human games and then on self-play. For life and death, the significance was that the neural networks learned shape and reading patterns from data rather than from hand-coded rules, and they evaluated positions in a way that captured the whole-board context—including ko threats and the strength of surrounding groups—that classical local analysis often abstracted away.
AlphaGo's successors, particularly AlphaGo Zero and AlphaZero, learned entirely from self-play and produced life-and-death judgments that sometimes contradicted classical theory. In some positions, sequences long considered dead were shown to be alive under optimal play, and vice versa. The most famous example is the three-three invasion and related corner sequences, where AI play revealed that classical joseki results were not always optimal—a finding that rippled into life-and-death analysis of corner groups.
The Present: Engine-Informed Theory and Open Questions
Today, professional and strong amateur players routinely consult Go engines—programs like KataGo and Leela Zero—when analyzing life-and-death positions. The engines provide not just a verdict but a variation tree showing the principal sequence and the evaluation of alternatives, which functions as a kind of automated deep reading. This has changed how players study: instead of memorizing problem solutions, they can query an engine and see why a position is alive or dead, including the ko threats and whole-board considerations that determine the result.
The open questions are genuine. First, the relationship between shape intuition and engine evaluation is not fully understood: engines sometimes prefer moves that violate classical shape principles, and players debate whether the classical principles were wrong or merely approximate. Second, ko and seki positions remain hard to evaluate even for engines, because their status depends on the whole board and on the order of play in a way that local search cannot fully capture. Third, the teaching of life and death is in transition: problem collections remain central, but the solutions in older collections are sometimes now known to be incomplete or wrong, and the field has not yet produced a comprehensive engine-verified replacement.
What has not changed is the underlying logic. A group lives if it can make two eyes; it dies if the attacker can prevent that; and the analysis of any particular position still requires reading, shape judgment, and attention to the whole board1. The history of the subfield is the history of building better tools for that analysis—from the problem collections of the Go houses to the neural networks of the present.
References
- Life and death (introduction, Examples, Eyes and false eyes, Status of groups of three, Status of groups with more than four empty points, Explanation, Seki), Wikipedia
- Shape (Go), Wikipedia
- List of Go terms (Seki, Tesuji), Wikipedia
- Computer Go, Wikipedia
- AlphaGo (History, introduction), Wikipedia
- AlphaGo versus Lee Sedol, Wikipedia
The frameworks of Life And Death
Each framework is a way of working the field has used: when it held, and the question it answers.
- 1600-Present
- 1700-Present
- 1700-Present
- 1700-Present
- 1950-Present
- 1970-2005
- 2006-2016
Monte Carlo Tree Search
How can random playouts estimate a position's value?
- 2016-Present
- 2017-Present
Engine-informed theory
What do the engines teach us about life and death?
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 Brute-Force Local Search
Monte Carlo tree search replaced exhaustive local brute-force search as the working method for computer Go, because random playouts gave usable position estimates where hand-crafted evaluation had failed. MCTS programs surpassed the earlier brute-force programs in tactical strength in the mid-2000s.
AlphaGo grew out of Monte Carlo Tree Search
AlphaGo kept Monte Carlo tree search as its search engine and added deep policy and value networks to guide it, so the 2016 system grew directly out of the MCTS approach. The playouts were replaced by neural evaluation while the tree search remained the core.
Carried by AlphaGo (2016).
Engine-informed theory grew out of AlphaGo
Engine-informed theory grew out of the AlphaGo and AlphaGo Zero results, as professionals analyzed the engines' life-and-death evaluations and revised classical judgments. The later theory is the human interpretation of what the AlphaGo line of programs showed.
Carried by AlphaGo Zero (2017).
Deep Reading (Yomi) influenced by Tsumego Problem Collections
The house tradition trained students by working through graded tsumego collections, and the deep reading (yomi) that professionals prized was built on this repeated problem study. The collections supplied the standard positions and sequences that readers learned to recognize and then to read out.
Carried by Gengen Gokyo.
Nakade and eye-space theory influenced by Tsumego Problem Collections
Classical problem collections encoded the results of generations of analysis of enclosed spaces, and the nakade principle was abstracted from these worked examples. The vital-point rule for reducing eye space is a compression of what the problem sets demonstrated.
Katachi Shape Theory influenced by Deep Reading (Yomi)
The katachi school systematized the selective reading that strong players used to prune the tree: shape knowledge told the reader which moves were worth reading. Fujisawa Shuko's writings on shape and the vital point turned the reader's pruning heuristics into teachable principles.
Carried by Fujisawa Shuko.
Brute-Force Local Search influenced by Deep Reading (Yomi)
Early computer Go programs tried to replace human deep reading with exhaustive local search, and their failure in tactical fights defined the problem the field had to solve. The brute-force tradition of yomi was the standard the programs were measured against and could not match.
Katachi Shape Theory influenced by Nakade and eye-space theory
Katachi theory generalized the nakade and eye-space results into a broader account of how the shape of surrounding stones determines life and death. Writers of the Japanese analytical school organized positions by shape type, building on the classical eye-space vocabulary.
Carried by Fujisawa Shuko; Ishida Yoshio.
AlphaGo influenced by Katachi Shape Theory
AlphaGo's networks were trained first on human games and learned shape and vital-point patterns that the katachi school had codified, and its play in life-and-death positions reflected that shape knowledge. The neural evaluation captured whole-board context that classical shape theory had abstracted away.
Carried by AlphaGo (2016).
Seki Theory influenced by Tsumego Problem Collections
Seki positions appeared among the special cases in classical problem collections, where mutual-life shapes were recorded alongside ordinary life-and-death problems. The later theory of seki organized and explained these collected cases.
Engine-informed theory influenced by Nakade and eye-space theory
Engine-informed theory revisited classical nakade and eye-space judgments, confirming some and overturning others, so the classical results remained the reference points the new theory revised. The vocabulary of eye space carried into the engine-era analysis.
Beyond the field
Where Life And Death's ideas came from in other fields, and what they led to there: the strongest 12 of 13 links.
Monte Carlo Tree Search grew out of Monte Carlo Methods in Reinforcement Learning
MCTS is a direct descendant of Monte Carlo methods: it applies random sampling to estimate the value of game positions, and UCT formalises this as a bandit-based Monte Carlo sampling scheme. The 2006 papers by Coulom and by Kocsis and Szepesvári made this derivation explicit.
Carried by Rémi Coulom; Levente Kocsis; Csaba Szepesvári.
AlphaGo led to Neural self-play in Chess
AlphaGo's later Zero version introduced pure neural self-play without human game data, and that method became the basis of neural self-play as a general framework. The supervised-policy AlphaGo of 2016 was the immediate predecessor from which the self-play variant was refined.
Carried by AlphaGo Zero (2017).
AlphaGo led to AlphaZero in Computer Chess
AlphaZero was built directly as a generalisation of AlphaGo Zero, dropping the human game records and Go-specific tuning to handle chess, shogi and Go with one algorithm. The AlphaGo line supplied the architecture, the self-play pipeline and the MCTS-plus-network evaluation that AlphaZero reused.
Carried by AlphaGo Zero (2017); David Silver.
AlphaGo influenced by Self-play in Game Playing Ai
AlphaGo trained its policy and value networks by self-play, a technique established in Samuel's 1959 checkers program and in TD-Gammon. The 2016 Nature paper describes reinforcement learning from games the system played against itself.
Carried by Arthur Samuel; Gerald Tesauro.
Brute-Force Local Search grew out of Heuristic search in Planning And Search
Brute-force local search in computer Go grew out of heuristic search techniques, applying them move-by-move without full tree expansion.
Brute-Force Local Search influenced by Adversarial game-tree search in Planning And Search
Adversarial game-tree search provided the minimax and evaluation framework that brute-force local search methods in Go adapted for move selection.
Brute-Force Local Search grew out of State-space search in Planning And Search
Brute-force local search is a specialized form of state-space search in which the state is the board position and only the local neighborhood is expanded.
Monte Carlo Tree Search influenced by Classical Game Tree Search in Game Playing Ai
MCTS grew out of classical game-tree search: its tree policy expands nodes and backs up values in the manner of minimax search, while replacing exhaustive evaluation with random rollouts. The 2006 Go papers by Coulom and by Kocsis and Szepesvári built directly on this search tradition.
Carried by Rémi Coulom; Levente Kocsis; Csaba Szepesvári.
Monte Carlo Tree Search replaced Alpha–beta pruning in Computer Go
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.
Monte Carlo Tree Search influenced by Adversarial game-tree search in Planning And Search
MCTS was developed within adversarial game-tree search, keeping the alternating-move tree of two-player games while replacing alpha-beta evaluation with Monte Carlo simulation. The 2006 Go work by Coulom and the UCT paper of Kocsis and Szepesvári sit in that lineage.
Carried by Rémi Coulom; Levente Kocsis.
AlphaGo influenced AI/engine-informed Go theory in Go
AI/engine-informed Go theory arose after AlphaGo's 2016 victory, when professional players studied the program's moves and revised opening and positional judgements. The later theory takes its questions and findings from AlphaGo's play.
Carried by AlphaGo (2016).
AlphaGo influenced by Neural-Network Engine Analysis in Backgammon
AlphaGo's use of a neural network to evaluate positions descends from TD-Gammon, which learned backgammon evaluation by neural network and self-play. The AlphaGo team cited this line of neural game evaluation as background.
Carried by Gerald Tesauro.

Life And Death is a planet in Nousfarer
Its city has a tower for each of these 9 frameworks. Land, walk in, and climb one concept at a time.
Land on Life And DeathRelated fields
Fields its ideas are linked to
- Planning And Search 4 links
- Game Playing Ai 2 links
- Computer Go 2 links
- Reinforcement Learning 1 link
- Chess 1 link
- Computer Chess 1 link
- Go 1 link
- Backgammon 1 link