不自然な計算結果

[nekobean]

以下の牌姿で7pの和了確率が高いのは違和感がある

東一局0本場 東家 3巡目 ドラ: 一筒
456678m02337p056s

version 0.9.5

## 場況
東一局0本場 東家 3巡目 ドラ: 一筒
456678m233r57pr556s
1向聴 (一般形: 1向聴 七対子: 3向聴 国士無双: 13向聴)

## 計算結果
打: 七筒, 有効牌: 2種8枚, 14p
期待値: 6383点, 和了確率: 39.30%, 聴牌確率: 94.43%
打: 五索, 有効牌: 3種12枚, 6p47s
期待値: 6333点, 和了確率: 38.65%, 聴牌確率: 93.75%
打: 六索, 有効牌: 5種16枚, 1346p5s
期待値: 6320点, 和了確率: 38.96%, 聴牌確率: 91.95%
打: 二筒, 有効牌: 5種16枚, 36p457s
期待値: 6256点, 和了確率: 38.67%, 聴牌確率: 94.30%

version 0.9.5 で向聴戻し、手変わりを OFF にすると version 0.9.0 と同じ順番になる

## 場況
東一局0本場 東家 3巡目 ドラ: 一筒
456678m233r57pr556s
1向聴 (一般形: 1向聴 七対子: 3向聴 国士無双: 13向聴)

## 計算結果
打: 六索, 有効牌: 5種16枚, 1346p5s
期待値: 5435点, 和了確率: 34.52%, 聴牌確率: 90.35%
打: 二筒, 有効牌: 5種16枚, 36p457s
期待値: 4509点, 和了確率: 31.39%, 聴牌確率: 90.35%
打: 三筒, 有効牌: 3種12枚, 146p
期待値: 4383点, 和了確率: 29.04%, 聴牌確率: 83.78%
打: 五索, 有効牌: 3種12枚, 6p47s
期待値: 4247点, 和了確率: 28.55%, 聴牌確率: 83.78%
打: 七筒, 有効牌: 2種8枚, 14p
期待値: 3769点, 和了確率: 24.05%, 聴牌確率: 73.79%

version 0.9.0

## 計算結果
打: 六索, 受入枚数: 5種16枚, 有効牌: 1346p5s
期待値: 5935点, 和了確率: 34.25%, 聴牌確率: 90.05%
打: 二筒, 受入枚数: 5種16枚, 有効牌: 36p457s
期待値: 5305点, 和了確率: 32.20%, 聴牌確率: 90.61%
打: 三筒, 受入枚数: 3種12枚, 有効牌: 146p
期待値: 5123点, 和了確率: 30.70%, 聴牌確率: 84.52%
打: 五索, 受入枚数: 3種12枚, 有効牌: 6p47s
期待値: 5114点, 和了確率: 30.55%, 聴牌確率: 84.89%
打: 七筒, 受入枚数: 2種8枚, 有効牌: 14p
期待値: 4568点, 和了確率: 27.02%, 聴牌確率: 74.79%

[Starry-Miracle]

Sorry I am not expertise in c++ programming nor in graph theory.

One of my guess is the result of tegawari is being count multiple times?
(Sorry I am not capable to understand the edge count logic in src/mahjong/core/expected_score_calculator.cpp)

I think one example make things really clear is
123m123p3334444z OR 123m123p33334444z in 15/16巡目
(where it is a forced sequence to cut 3z4z for any tile to tenpai. I expect 'kan' is not an option?)

The difference in win probability between tegawari on/off option in v0.9.5 comes from nowhere.

[nekobean]

Thank you for your interest in this repository.

As you pointed out, I suspect that hand transitions are being excessively explored in certain situations.
The following parts might be problem.

---

向聴戻し(discards that increase the shanten number) and 手変わり(improvment tile) are only explored if the following conditions are met

d(h, origin) + shanten(h) < shanten(origin) + 1

d(h, origin): the number of tile exchanges required to transform from the original hand to the current hand
shanten(h): shanten number of hand h

mahjong-cpp/src/mahjong/core/expected_score_calculator.cpp

Line 311 in 482e760

distance(hand_counts, hand_org) + shanten < shanten_org + config.extra;

mahjong-cpp/src/mahjong/core/expected_score_calculator.cpp

Line 380 in 482e760

distance(hand_counts, hand_org) + shanten < shanten_org + config.extra;

• If 手変わり option enabled and the above condition are met, transitions are searched for when drawing any tile.
  Otherwise, transitions are searched for only when drawing necesarry tiles (tiles that decrease the shanten number).

• If 向聴戻し option enabled and the above condition are met, transitions are searched for when discarding any tile in hand.
  Otherwise, transitions are searched for only when discarding tiles that remain the shanten number.