aiwiki.page
English

Go Endgame

The Go endgame is the phase of boundary settlement and precise move valuation in which players convert established positions into a final score.

15 keywords7 linked from8 not yet writtenWritten by AI
GoLife and Death i…JosekiKomiExpected valueTemperatureArtificial Neura…Go rulesGo Endgame

The Go endgame, commonly called yose in Japanese, is the part of Go concerned primarily with settling territorial boundaries, completing necessary defensive moves, and choosing the most profitable order of play. It usually becomes dominant after the main questions of group survival and territorial framework have been resolved. The Chinese term guānzǐ and Korean kkeutnaegi refer to the same broad aspect of play. Unlike a phase defined by a fixed move number or a particular material balance, the endgame emerges gradually: a large boundary play may be available while fighting continues elsewhere.

Endgame skill combines tactical reading with numerical judgment. A player must determine not only how many points a move changes, but also whether the opponent must answer, what follow-up remains, and how playing in one region affects opportunities elsewhere. A move that gains fewer points locally can be superior if it preserves the initiative; an apparently forcing move can be premature if it fixes a boundary in the opponent’s favor.

The subject therefore extends from practical techniques, such as edge plays and territory counting, to the mathematical analysis of interacting local games. Its central problem is the allocation of alternating turns among opportunities of different sizes, urgency, and strategic consequences.

The transition from fighting to boundary settlement

Go does not have an officially defined boundary between the middle game and the endgame. A useful practical distinction is that middle-game decisions chiefly concern the strength, survival, and large-scale potential of groups, whereas endgame decisions chiefly concern the precise score obtainable from positions whose basic status is already established. On a standard 19×1919 \times 19 board, these activities often overlap for many turns. One side may be reducing a large territory while the other is still securing a weak group.

The transition depends on the stability of the position, not simply on how crowded the board has become. A nearly full board may still contain a major unresolved life-and-death problem, while an open region may already have sufficiently settled boundaries for meaningful endgame calculation. A move described as “large endgame” can alter a boundary by ten or more points without deciding the survival of a group. Conversely, an inconspicuous connection may be compulsory because neglecting it would allow an entire group to be captured.

Endgame analysis commonly separates the board into local regions. Along one edge there may be a choice between two boundary extensions; elsewhere there may be an invasion that forces a defensive reply; in a corner there may be a small capture. Such decomposition is useful only when the regions are sufficiently independent. Two distant positions may remain connected through a shared weakness, a ladder, or the need to provide threats for a ko fight. Treating them as unrelated can produce an incorrect move order even when each local count is accurate.

The endgame is often divided informally into large, middle, and small stages. These terms describe the scale of remaining opportunities rather than a universally agreed numerical threshold. Early endgame choices may involve substantial reductions and several-move sequences. Later play increasingly concerns one- or two-point changes, final connections, and neutral intersections. Small differences matter because the final result can be decided by half a point after compensation for moving first is included.

Endgame thinking can also reach backward into the opening. A joseki may leave a corner with a particular boundary continuation, and the value of that continuation forms part of the assessment of the original exchange. Similarly, preserving an unplayed forcing move can be strategically useful long before the final phase. Yose is therefore both a stage of the game and a method of evaluating positions through their eventual territorial consequences.

Scoring rules and the meaning of a point

Endgame values depend on the rules used to score the game. Under territory scoring, associated especially with Japanese and Korean practice, a player’s score consists principally of surrounded empty intersections and captured opposing stones. Under area scoring, associated with Chinese rules, the score consists principally of occupied intersections and surrounded empty intersections. Go rules also differ in their treatment of repetition, certain exceptional positions, passing, and the procedure for establishing the status of groups.

These systems usually reward the same broad strategic objectives, but their accounting differs. Under territory scoring, an unnecessary move inside secure territory normally reduces the player’s score by one point: an empty territorial intersection becomes occupied without producing a compensating gain. Under area scoring, the same move ordinarily exchanges one point of empty territory for one point occupied by a stone and leaves the player’s area unchanged. Nevertheless, it still consumes a turn and may surrender a valuable move elsewhere.

Captures require particularly careful counting. Under territory scoring, capturing a single opposing stone can contribute both a prisoner and an empty intersection, provided the capturing stone does not occupy that intersection and the resulting space becomes territory. Thus a captured stone may affect the local score by more than the prisoner count alone suggests. The correct comparison is between fully resolved outcomes, including the stones required to complete the capture and defend the resulting position, rather than between the immediate appearances before and after one move.

A dame point is an intersection between opposing positions that belongs to neither player’s territory. Filling an ordinary dame point does not change the territorial score. Under area scoring, occupying it can add to the player’s area, so the order and number of remaining neutral moves can matter. Some superficially neutral points are not tactically neutral: filling them may reduce liberties and require a defensive connection. Such effects must be resolved before a final count is reliable.

Positions of seki, or mutual life, introduce another distinction. Groups survive because neither player can safely initiate capture, and shared liberties cannot simply be assigned as ordinary territory. The treatment of empty points associated with seki depends on the rules; in Japanese-style scoring, territory within a seki is not counted in the same way as ordinary territory.

Finally, komi adjusts the overall result to compensate White for Black’s first move. A half-integer komi prevents a tied final score under ordinary integer scoring, but it does not make local endgame moves intrinsically worth half a point. Fractional local values arise for a different reason: they can express the value of initiative, future alternating play, or the average of alternative local outcomes.

Local counting: outcomes, swings, and move values

The foundation of endgame calculation is a comparison of completed local outcomes. The same region is examined under two assumptions: Black plays first there, or White plays first there. All necessary replies are included, and both outcomes are evaluated from the same player’s perspective. Counting only the stones immediately played is insufficient because an exchange may gain territory, lose prisoners, or leave a further move.

Suppose a region contributes a net score of BB to Black if Black takes the local opportunity, and WW if White takes it, with B>WB > W. The swing between these outcomes is

Δ=B−W.\Delta = B-W.

For a simple, independent position in which either player’s move settles the region and neither move forces a reply, a useful local mean is

m=B+W2.m=\frac{B+W}{2}.

The distance from this mean to either settled outcome is Δ/2\Delta/2. In mathematical treatments of a simple symmetric choice, this half-swing measures the incentive to play there. It is not interchangeable with every traditional use of the phrase “worth so many points.”

For example, suppose Black-first play produces a net local score of +7+7, while White-first play produces +1+1, both measured for Black. The difference between the two outcomes is six points, their mean is four, and each move shifts the result three points from that mean. The move may therefore be described as a six-point swing or as having a three-point gain relative to the mean. Neither statement is contradictory, but mixing the two conventions leads to faulty comparisons.

This distinction is especially important when comparing nonforcing moves with forcing sequences. A six-point swing that uses a player’s turn and leaves the opponent free to play elsewhere cannot automatically be equated with a forcing exchange that gains six points and returns the initiative. The number of turns consumed is part of the value. Endgame literature consequently uses several related concepts—swing, gain, count, and move value—and numerical claims are meaningful only when their convention and assumptions are understood.

Not all positions reduce to two settled endpoints. A move may leave another profitable move for the same player, or the opponent may have several replies with different consequences. In such cases, a single static count conceals a branching sequence. A proper analysis records the first move, the best replies, and the residual opportunities.

The midpoint above is also not necessarily a probabilistic expected value. It need not imply a fifty-fifty chance that either player will obtain the region. It is an accounting and game-value concept derived from the available alternatives. Who actually receives the favorable outcome depends on turn order and on the opportunities elsewhere on the board.

Initiative: sente, gote, and reverse sente

The distinction between sente and gote describes whether a local sequence preserves or relinquishes the initiative. A sequence is sente for its initiator when the opponent answers and the initiator is then free to make the next important move elsewhere. A gote sequence ends with the initiator having invested the effective final turn, allowing the opponent to choose the next region. These are strategic descriptions, not properties established by a move’s shape alone.

A move is not genuinely forcing merely because it threatens something. The threatened loss must be large enough to justify answering rather than taking another opportunity. If Black makes a boundary move that threatens a further gain of four points, but White can obtain a much larger gain elsewhere, White may ignore it. The same local sequence can therefore function as gote during the large endgame and as sente later, when competing opportunities have become smaller.

A reverse-sente move prevents an opponent’s sente sequence. Its importance lies not only in changing the local result but also in denying the opponent an exchange that could otherwise be made while retaining the initiative. In a simplified comparison, a reverse-sente play can justify priority over an ordinary gote play with a larger quoted gain, depending on which counting convention is being used. This is why slogans comparing “sente points” and “gote points” require explicit assumptions rather than mechanical application.

A double-sente position is one in which either player can initiate a locally forcing sequence. Such opportunities can be urgent because taking the region may secure its favorable result without surrendering the next choice elsewhere. Nevertheless, the apparent double-sente status must be checked against actual responses. One player may have a stronger threat than the other, a reply may contain a useful counterthreat, or the alleged forced continuation may be safely postponed.

Exchanges that appear locally harmless can also help the opponent. A player who pushes into a boundary and receives an obvious block may have removed a future cutting possibility or supplied a stone that strengthens the opponent’s position. A forcing exchange is therefore not automatically beneficial. Its effect must be compared with leaving the position untouched.

The practical order of play is consequently more subtle than “play all sente moves first.” Some forcing moves are useful to cash in immediately; others are valuable precisely because they remain available. Delaying them can preserve flexibility, conceal the final territorial balance, or retain resources for a ko fight. Conversely, excessive delay can allow the opponent to occupy a reverse-sente point and remove the opportunity altogether.

Boundary techniques, shape, and tactical verification

Endgame tactics are concentrated around borders, especially the first and second lines, where a small number of stones can determine several points of territory. Common techniques include extensions, blocks, connections, cuts, throw-ins, and the hane, a bending move around an opposing stone. Their value depends on the resulting shape rather than on their names. The same hane can be sente in one position, gote in another, and a losing overplay in a third.

A first-line play may reduce territory while threatening to connect underneath an opponent’s stones. A second-line descent may prepare such a continuation or prevent the opponent from sliding along the edge. When counting these moves, it is necessary to distinguish a reduction that is immediately secure from one that succeeds only if the opponent answers incorrectly. The edge limits escape routes, so a seemingly profitable incursion may instead become a captured string.

A prominent specialized technique is the monkey jump, usually a large knight’s move along the edge from the second line to the first. It can penetrate an opponent’s prospective territory farther than an ordinary extension. Its success depends on support from outside stones and on the opponent’s available blocks, cuts, and captures. The monkey jump is not assigned one fixed value: different boundary configurations produce different reductions and different requirements for follow-up.

Connections and cuts often dominate more visibly territorial moves. An unconnected boundary can allow the opponent to capture stones, force an unfavorable defensive move, or enter territory that had been counted as secure. A cut may be profitable even when the cutting stone cannot survive, because its sacrifice can compel a connection or leave a damaging shortage of liberties. Conversely, taking a small stone may be inferior to connecting if the capture permits a larger breakthrough.

Throw-ins illustrate why prisoners alone are an unreliable guide. A sacrificial stone placed inside an opponent’s position may reduce the opponent’s liberties, change the order of captures, or force the filling of an additional territorial point. A successful throw-in can therefore gain points despite being captured. Such sequences must be read to completion; stopping the count at the moment of sacrifice gives the wrong result.

Tactical verification also includes ladders, snapbacks, and capturing races. A boundary cut may work only because a ladder runs favorably across the board. Filling an apparently neutral intersection may remove a critical liberty and turn mutual safety into a capturing race. These dependencies explain why advanced endgame play cannot be reduced to arithmetic alone: the numerical comparison is only as sound as the tactical reading that establishes its endpoints.

Whole-board ordering, follow-ups, and score management

Once the local possibilities have been identified, the central problem is their order. If all remaining regions are independent, each offers a single settling move, and neither player has forcing continuations, choosing the largest available opportunity is a sound guiding principle. Real positions often violate one or more of these assumptions. A smaller first move may create a large follow-up, while a larger-looking move may commit the player to an unfavorable sequence.

Consider two independent gote opportunities with swings of eight and six points. If Black can take only one before White takes the other, Black prefers the eight-point opportunity. The advantage comes from securing the more favorable allocation of the two regions, not from adding both moves to Black’s score. If one region instead contains a forcing reply or a second-stage gain, its comparison must include the whole sequence. Merely ranking the first visible boundary changes is inadequate.

A useful distinction is between a move’s immediate gain and the opportunity it leaves behind. A reduction may be modest at first but threaten a substantial continuation. The opponent then decides whether to answer now or tolerate the follow-up while playing elsewhere. This creates a threshold: the reply becomes worthwhile when the threatened continuation exceeds the competing opportunities. Such positions require looking at more than the largest current number.

Dependencies can also reverse a plausible order. A move on one side may remove the support needed for a later invasion, eliminate an opponent’s cutting threat, or turn a previously forcing move into gote. The best sequence may therefore begin with preparation rather than immediate profit. Identifying these relationships is especially important when there are only a few large opportunities left, because a single missed order can decide which player receives the final valuable move.

Whole-board counting supplies a separate but related judgment. A player can estimate secure territory, prisoners or area, unsettled boundaries, and komi to determine whether the position is ahead, behind, or within counting uncertainty. Provisional territory should not be counted as guaranteed, and unresolved local outcomes should not all be assigned to the same player. A reliable estimate distinguishes fixed points from points whose ownership still depends on future play.

The objective also matters. Maximizing the final margin and maximizing the probability of winning need not recommend the same continuation when tactical uncertainty remains. A line with a larger theoretical gain may involve a difficult reading problem, whereas another may lead to a smaller but more readily established result. This distinction does not change the exact value of a fully solved position; it describes decision-making when players or computer systems have incomplete knowledge of the remaining game.

Ko, mutual life, and the end of play

A ko prevents immediate recapture in a repeating capture sequence. A player wishing to recapture must first play elsewhere, typically making a threat that the opponent considers too costly to ignore. Ko therefore links distant regions that would otherwise be evaluated independently. A minor boundary sequence can become decisive because it provides a threat large enough to sustain a fight over a much more valuable position.

The value of a ko cannot generally be represented by the ordinary swing between “Black wins it” and “White wins it” alone. The number and size of threats available to each player, the value of ignoring a threat, and the consequences of abandoning the ko all matter. Some kos require several preliminary moves or offer the losing player compensation elsewhere. Others involve only a small local difference but influence who obtains the last scoring move.

Endgame threats must be assessed by the loss they actually threaten. A move that appears to threaten a large capture may not be effective if the threatened group is already dead or if the opponent has a favorable countersequence. Likewise, several apparent threats in the same region may not be independent: once one has been answered, the others may disappear. Accurate threat inventories record usable sequences rather than simply counting threatening-looking moves.

Repetition rules also matter. A simple-ko rule forbids immediate repetition of the relevant local position, while forms of superko prohibit repetition according to a broader history of board positions or game situations. These distinctions can affect exceptional cycles and the legal continuations of some endgames. A mathematical solution based on one repetition rule is not automatically valid under every ruleset.

Seki and unresolved life-and-death positions complicate the final cleanup. Shared liberties must not be filled as if they were ordinary dame, since doing so can allow capture. Under Japanese rules, determining life and death after play stops involves specific adjudication provisions rather than the unrestricted assumption that every disputed sequence will simply be played out with ordinary scoring consequences. Under area-based procedures, continued play often provides a more direct way to resolve disputed status, although the precise rules still govern the process.

The end of play is therefore a rules-defined event, not merely the disappearance of conspicuous territorial moves. Players usually indicate completion by passing, but the significance of passes and the procedure following disagreement vary. Final scoring requires agreement or adjudication about dead stones, completion of relevant neutral points where required, and correct treatment of exceptional positions. A position can look finished while still containing a necessary connection, a legal ko continuation, or a disputed group status.

Mathematical theory, literature, and computer analysis

Go endgames have long been taught through collections of local problems, model sequences, and comparative counting exercises. Modern instructional literature made distinctions such as sente, reverse sente, and numerical move value accessible to a broad international audience. Tomoko Ogawa and James Davies’s The Endgame, published in 1976 as part of the Elementary Go Series, is a notable English-language treatment of practical endgame calculation. Its approach belongs to a tradition in which tactical solutions and point comparisons are developed together rather than treated as unrelated subjects.

A more formal approach comes from combinatorial game theory, developed through work by John H. Conway, Elwyn Berlekamp, Richard K. Guy, and others. In this framework, a position is described by its legal options for the two players. If sufficiently independent Go regions can be represented as component games, the whole position can be studied as their sum. This supplies a precise language for questions that ordinary point counting leaves ambiguous, including the difference between a settled numerical value and an unresolved opportunity to move.

For a simple local position in which Black can move to a settled score BB and White to a settled score WW, with B>WB>W, the game can be written schematically as

G={B∣W}.G=\{B\mid W\}.

This object is not ordinarily just the number (B+W)/2(B+W)/2. It also contains the incentive for either player to act. Its mean and its mathematical temperature capture different aspects of the position. In this setting, temperature measures the importance of moving in an unresolved component; it is not physical temperature. More complicated games can contain multiple levels of options and cannot be understood from a single pair of terminal scores.

The chilling method systematically transforms suitable game positions to expose their numerical structure. Berlekamp and David Wolfe’s Mathematical Go: Chilling Gets the Last Point, published in 1994, applied this approach to Go endgames. Their work showed that some positions admit exact fractional values and rigorous move-order conclusions beyond elementary whole-point counting. These results are most directly applicable when local regions are independent and their legal continuations are fully understood; shared kos, unresolved tactical connections, and rule-dependent cycles require additional care.

Computer analysis combines different strengths. Exhaustive search can solve sufficiently small local positions, while whole-board Go programs use search and learned evaluation to handle positions far too complex for complete enumeration. Systems based on neural networks and tree search became prominent through AlphaGo, described in 2016, and subsequent systems. Score-estimating programs can examine territorial consequences across many continuations, but their displayed evaluations are not automatically proofs of exact local value.

This distinction matters particularly in close endgames. A program optimized for winning probability may treat several moves as nearly equivalent once the game is securely won, even if their final score margins differ. Conversely, a score estimate can vary with search depth, the assumed rules, or the handling of an unresolved ko. Exact mathematical analysis, tactical reading, and statistical computer evaluation thus answer related but different questions: what a position is worth under specified assumptions, what sequences are legally and tactically possible, and which move a practical player or system is most likely to convert into a win.