Loading field map
Every shogi player confronts a fundamental tension: the king must be sheltered, but the pieces that build that shelter are also the pieces needed for attack. A castle that is too solid may leave its owner with no counterplay; a castle that is too loose may collapse under a single well-placed drop. Over the past century, castle theory has been organized around four successive frameworks, each offering a different answer to the question of how much defense is enough and what kind of defense best supports the overall strategy.
For most of the twentieth century, castle theory was inseparable from opening choice. Players aligned themselves with one of two grand traditions: the Ranging Rook Castle School or the Static Rook Castle School. These were not merely sets of formations but entire philosophies about the relationship between king safety and piece activity.
The Ranging Rook Castle School grew out of the Ranging Rook opening family, in which the rook slides to the center or left side of the board. Because the rook leaves its original post, the king can be moved into the space it vacated, allowing a compact, deep castle on the right side. The signature formation of this school is Anaguma (the "bear in the hole"), a fortress that tucks the king into the far corner behind three layers of gold and silver generals. Anaguma is extraordinarily difficult to break, but it consumes many pieces that might otherwise participate in an attack. The Ranging Rook Castle School therefore accepted a division of labor: the castle would absorb the opponent's assault while the ranged rook and a few supporting pieces generated counterplay on the other side of the board. Endurance, not speed, was the virtue.
The Static Rook Castle School, by contrast, emerged from Static Rook openings, where the rook stays on its original file and the king must find safety elsewhere—usually on the left side. Because the rook remains close to the action, the castle must be built quickly and must not consume pieces needed for a direct attack. The signature formations of this school are Yagura (a layered gold-and-silver structure that can transition into offense) and Mino Castle (a compact, resilient formation that leaves the rook free to operate). Where the Ranging Rook school separated defense from offense, the Static Rook school aimed to integrate them: the same generals that protect the king can also advance to support a pawn storm or a bishop exchange. This integration made Static Rook castles more flexible but also more vulnerable to prolonged sieges.
The two schools coexisted for decades, and their rivalry shaped how players trained and how tournaments were prepared. A professional player in the 1950s or 1960s typically specialized in one school, mastering its castle repertoire and learning to exploit the weaknesses of the other. Match strategy often revolved around forcing the opponent into a castle type they were less comfortable defending. The schools were not monolithic—within each, sub-variations proliferated—but the fundamental divide between compact endurance (Ranging Rook) and integrated flexibility (Static Rook) remained the organizing axis of castle theory until the 1980s.
Around 1980, a new generation of analysts began to question the school-based approach. Instead of asking "Which castle belongs to my opening?", they asked "Which castle gives the best position in this concrete situation?" This shift marked the emergence of Modern Analytical Castle Theory, a framework that treated castle choice as a context-dependent decision rather than a matter of strategic allegiance.
The key innovation was positional evaluation: analysts began to assign numerical or qualitative scores to castle formations based on factors such as king safety, piece mobility, pawn structure, and the timing of attacks. This was not yet computer-driven—the analysis was done by hand, using game records and deep study—but it was far more systematic than the intuitive judgments of the classical schools. Modern Analytical Castle Theory introduced hybrid formations that would have been unthinkable under the old school system. The Fortress Castle (a Static Rook formation that uses a gold and a silver to create a nearly impenetrable wall) became popular even in Ranging Rook games, and the Ishida Castle (a Ranging Rook formation adapted for rapid attack) was used in Static Rook contexts. The framework also revived and refined older castles that had fallen out of favor, such as the Nakahara Variation of Yagura, by showing that their defensive weaknesses were offset by specific attacking opportunities.
Early computer tools played a supporting role in this period. By the late 1990s, shogi engines could evaluate positions faster than any human, and analysts began using them to test castle variations that were too complex to calculate by hand. The engines did not yet dictate castle theory—their evaluations were often unreliable in the opening and middlegame—but they provided a new source of data. Modern Analytical Castle Theory absorbed this data selectively, using it to confirm or challenge human judgments rather than to replace them. The framework remained human-centered: its goal was to give the player a principled method for choosing a castle, not to find the single objectively best move.
After 2010, the rapid improvement of shogi engines—especially those using deep learning and massive self-play—fundamentally changed the relationship between human analysis and castle theory. AI-Driven Castle Optimization does not ask which castle is most elegant or which castle fits a strategic plan. It asks only one question: which king position maximizes the engine's calculated winning probability?
The answers have been startling. Classical fortresses like Anaguma, once considered nearly invincible, are now often avoided in top-level play because engines show that the pieces invested in the castle could be better used for attack or for controlling key squares. Minimalist king positions—sometimes just a single step to the side, with no formal castle structure at all—are preferred when the engine judges that the king is safe enough and the extra piece can generate decisive threats. The AI-driven framework does not reject the classical schools outright; it re-evaluates every formation on a case-by-case basis, and some classical castles (such as Mino) remain viable in specific positions. But the principle has shifted: defense is no longer a fixed commitment but a dynamic trade-off that must be recalculated every move.
Modern Analytical Castle Theory has not been absorbed into AI-Driven Castle Optimization; the two frameworks coexist, but with a clear division of labor. Modern Analytical methods are still used by human players for opening preparation and for understanding the strategic ideas behind a castle. AI-Driven Optimization, meanwhile, provides the final arbiter of correctness: when a human-prepared castle line is tested against an engine, the engine's verdict often overrides traditional positional judgment. In professional shogi, most top players now use engines during preparation to check their castle choices, and many have adopted AI-recommended king positions that would have seemed reckless a generation ago.
Today, the leading frameworks are Modern Analytical Castle Theory and AI-Driven Castle Optimization. They agree on one fundamental point: castle choice should be determined by the concrete demands of the position, not by school allegiance. They disagree on how to determine those demands. Modern Analytical Castle Theory relies on human-readable positional principles—king safety, piece coordination, pawn structure—that can be taught and discussed. AI-Driven Castle Optimization relies on engine evaluations that are often opaque to human intuition; a player may know that a certain king position is best without fully understanding why.
The classical schools remain important as pedagogical foundations. Every serious shogi student still learns Anaguma, Yagura, and Mino, because these formations teach the core concepts of king safety and piece economy. But the classical schools no longer dictate tournament strategy. A professional player today might use a Ranging Rook opening with a Static Rook castle, or a Static Rook opening with a minimalist king position, depending on what the engine recommends. Castle theory has moved from a world of competing traditions to a world of continuous optimization, where the only authority is the win probability on the board.