How Balance Shogi came to be
I trained an AI for Fuseki Shogi and Sente won 60% of the time. This is how I stopped hunting for a magic number in the rules and let the players balance the game themselves.
Balance Shogi itself is playable on this site; start from the Balance Shogi page.
1. Sente wins too much
Fuseki Shogi is a variant in which both sides place 20 pieces on an empty board, one at a time, before anything moves. Instead of a single starting position there are about 1037 of them. I wanted a shogi that does not feel researched to death, so I built an AI you can play in the browser: a network that places the pieces, trained by self-play, and YaneuraOu for everything from move 41 on.
As the network improved, one thing became impossible to ignore. Sente won too much, and more training did not bring it down. Nor, it turned out, is this special to Fuseki Shogi.
Within the same tournament the number rises from the qualifiers to the final. The advantage grows with the strength of whoever is playing, and cutting Sente's clock did not close it.
2. The rule text cannot fix it
My first instinct was to fix it in the text of the rules: restrict the first move to an edge pawn, forbid placements where pieces attack each other, reduce the number of ranks. The only change that really helped was a geometry close to ordinary shogi — place within three ranks, pawns on the third — and even that left Sente ahead.
| Rule tried | Sente wins | Notes |
|---|---|---|
| Current (four ranks, 20 pieces) | 62.1% | trained placement network |
| No piece may attack another | 62.4% | no effect |
| Three ranks, 15 pieces | 59.4% | random placement |
| Three ranks, pawns on the third | 53.8% | random placement; the best of them |
Removing contact or changing the ranks barely touches the core of the advantage. What is left is the value of a tempo, and that depends on the strength of whoever is playing, not on the geometry. A rule that looks even at one strength will be off at another.
Chess also favours White, but at the top the draw rate rises and absorbs the advantage. Shogi has pieces in hand and almost no draws, so that escape route is closed. Even if I found a rule that measured 50% in my environment, nothing guarantees the same number for the newest engines. A fixed rule cannot follow strength. I stopped looking for a number.
Awkwardly, going deeper with the same engine pushed ordinary shogi down: Sente scored 54.0% at 200k nodes and 50.1% at 1M. That suggests the rising tournament figures owe more to opening books and modern evaluations than to depth alone. Either way, a rule calibrated here carries no guarantee over there.
3. The placer and the chooser
Only one mechanism is known to stay balanced regardless of strength: one player divides, the other picks first. It is how two people share a cake, and it is standard in Hex and in Gomoku. Gomoku's Swap2 opening — the first player puts down two black stones and one white, the second player takes a colour — brought the first player's 67% down to about 52%.
Applied naively to Fuseki Shogi it breaks. If a player may pick a camp after the placement phase, they simply build their own camp badly and take the opponent's. The one who divides is also the one who picks. That is why I had rejected the pie rule earlier.
This time I added a single step at the start of the placement phase, which splits the job three ways: place the kings, choose a side, build the camp.
The placer is neutral because they do not know their side; the chooser cannot sandbag because they have not built anything. Whatever advantage Sente still has is cancelled every game by the placer putting the Sente king on a slightly worse square.
One pan of the scale carries the first move, the other carries king safety, and the placer has no choice but to level them. Hence Balance Shogi.
4. Are the kings enough?
Whether the kings alone suffice depends on whether king squares carry enough value to cancel the first-move advantage. That can be measured. For every pair of squares — 36 for the Sente king by 36 for the Gote king — I let the placement network finish the placement and had the engine score the move-41 position.
king_pairs_iter2455_games.json.A king on the front rank, closest to the opponent, is a heavy loss; the corner of the back rank is best. The gap between the best and worst pair is 95 points, several times the first-move advantage. The kings are enough, with room to spare. The site's AI reads this table both when placing and when choosing.
5. How do you measure balance?
Here I was wrong for a long time. I was looking at the placer's win rate, and treating a number near 50% as balanced.
That measures nothing. The placer's score moves with the quality of the opponent's choice: against a poor chooser the placer comes out ahead, against a good one behind. In fact the table whose placer win rate looked closest to even turned out to be the most exploitable of the three I tried.
The right question is how much a chooser playing perfectly can take. Handed one pair of king squares, there is really only one thing for the chooser to do.
Whatever the gap is, the chooser collects it. Nothing the placer does afterwards changes that distance — and balance means making it small. The placer's win rate is no way to measure it.
6. Fixing the yardstick
Measured that way, the pairs the placer was handing over were not balanced at all.
The fault lay not in the geometry of the kings but in the yardstick the table used. I was scoring the move-41 position with the engine's evaluation and converting it to a win rate through a formula that had no justification behind it. Calibrated against actual results, the conversion had been reading Sente far too high. That fixed the tilt.
The scatter between pairs remained. Scoring is a proxy; it never looks at balance directly. So I dropped the scoring and put tens of thousands of real games into the table instead, playing the candidates out and re-picking the 48 pairs closest to even.
What the chooser can take is exactly how far the dots sit from the dashed line. Halve the scatter and you halve the take.
The re-picked 48 were measured on the same games that selected them, so 3.1pt is an optimistic bound: select on a sample and score on the same sample and everything looks more balanced than it is. Confirmation on fresh games is under way.
7. The placer cannot win
This mechanism has a ceiling that no measurement can move.
A game that is finite, has no chance, and shows both players everything is determined (Zermelo's theorem, 1913): either the first player can force a win, or the second can, or both can force a draw. We may not know which, but we know it is one of them.
Fix one pair of king squares and what follows is an ordinary perfect-information game, so that pair has one of those three values. The chooser reads the value and takes the winning side — Sente if the pair is a Sente win, Gote if it is a Gote win. All the placer can do is look for a drawn pair. There is no path to a win anywhere.
So the value of perfectly solved Balance Shogi is either a draw or a win for the chooser; a win for the placer cannot happen. Hex, which has the same mechanism and no draws, is settled: the chooser wins, proven. Shogi keeps repetition and impasse, and that is where the possibility of a genuine draw survives.
This is not a design failure but something no rule can escape. Ordinary shogi is in the same position: if Sente can force a win, the toss decides the game. What you can aim at is not erasing the value but making the difference invisible at the strengths people actually play at. The whole job of Balance Shogi is to shrink the tilt of the pairs the placer hands over until neither a human nor an engine can see it. The precision of the table is the ceiling on fairness.
The strongest engines show a larger first-move advantage still. If the kings stop being enough, there is room to widen what gets placed from the kings to the kings and rooks.
8. Adding Nihikyō
On 13 September 2026 Balance Shogi gained one more placement restriction: each file of your camp may hold at most one of your rook and lances. I called it Nihikyō. Fuseki Shogi itself is unchanged.
The trigger was the placement AI's own games. Stacking the rook and both lances on one file — a “triple rocket” — had become a dominant strategy. In the newest network 55.7% of camps built it, and in games where the opponent built nothing, the side that did scored 11.4 points higher. It wins whether or not the opponent builds one too, so more training never wore it out.
The core of its strength is the two lances: lance and lance without the rook does nearly as well. But banning only doubled lances looked likely to push the AI to rook and lance, which would keep some of the value, so the rule counts the rook as well and the change only has to be made once. Shapes that can be answered by preparing a defence, such as pieces facing each other, are left alone. The only thing removed is stacking pieces that run along the same file.
The king table and every measurement above were taken under the rules before Nihikyō and have not been redone. The placement AI, too, is still the one trained under the old rules; it simply skips the moves that are now illegal.
Sources
- Sente now wins 94% in shogi AI tournaments (YaneuraOu, in Japanese)
- How large is the first-move advantage in shogi AI today? (YaneuraOu, in Japanese)
- Win rates and openings at the 36th World Computer Shogi Championship (in Japanese)
- Pie rule, Swap opening rule (Wikipedia)
- First-move advantage in chess (Wikipedia)
- Zermelo's theorem, Hex (Wikipedia)