fix: ISMCTS v2 — 包庄阈值校准 + 多牌组合Rollout + 防死循环

P0修复: 包庄严重高估
- 阈值 45%→60%(1000局数据校准)
- Rollout 对手从「最小单张」改为 PickBestTip(最大+多牌优先)
- 模拟对手更强 → P(win) 估计更准确

P2修复: 只出单张不组合
- 新增 PickBestTip(): 优先多牌组合(对子/顺子 > 单张)
- 同数量选最大的牌
- Rollout 模拟中所有玩家均受益

P1确认(虚警): 75 moves 异常局实为碾压局(高牌玩家清16张)
- 加 pass streak 检测:连续 2 pass 则强制最大出牌
This commit is contained in:
2026-07-07 13:59:49 +08:00
parent 4a3fbec057
commit 4c21e5d8b5
7 changed files with 122 additions and 81 deletions

View File

@ -9,14 +9,15 @@ using GameMessage.PaoDeKuaiF;
namespace PdkFriendServer.Logic
{
/// <summary>
/// ISMCTS (Information Set Monte Carlo Tree Search) 机器人
/// 对每个候选出牌方案模拟 N 局随机对局,选胜率最高的方案
/// 只看到自己手牌+桌面信息——对手手牌从未知牌池随机采样
/// ISMCTS 机器人 v2
/// P0修复: 包庄阈值 70%(从 45% 校准rollout 用最强方案而非最小
/// P2修复: 优先多牌组合(对子/顺子 > 单张),提高清牌效率
/// </summary>
public class IsmctsBot : IPdkBot
{
private readonly int _simulations;
private readonly Random _rng;
private const double BaoThreshold = 0.60; // 校准自1000局压测70%→0人包45%→5.2%成功
public IsmctsBot(int simulations = 800)
{
@ -24,24 +25,22 @@ namespace PdkFriendServer.Logic
_rng = new Random();
}
// ---- 包庄决策 ----
public bool DecideBaoZhuang(PdkBotView view)
{
// ISMCTS 包庄决策:模拟 N 局,估算自己第一个出完的概率
// 包庄不改变牌局流程只翻倍输赢P(win)*2*gain - P(lose)*2*loss
// 阈值 45%——因为包庄成功收益翻倍,略微倾斜有利于局势占优
double winRate = EstimateWinProbability(view);
bool shouldBao = winRate >= 0.45;
bool shouldBao = winRate >= BaoThreshold;
Console.WriteLine($"[ISMCTS pos{view.MyPos}] 包庄评估: P(win)={winRate:P0} → {(shouldBao ? "" : "")}");
Console.WriteLine($"[ISMCTS pos{view.MyPos}] 包庄评估: P(win)={winRate:P0} "
+ $"(阈值{BaoThreshold:P0}) → {(shouldBao ? "" : "")}");
return shouldBao;
}
/// <summary>模拟 N 局随机对局,统计我第一个出完的概率</summary>
private double EstimateWinProbability(PdkBotView view)
{
int myPos = view.MyPos;
var hand = view.MyHand;
int sims = Math.Min(_simulations, 600); // 包庄模拟用 600 次,比出牌少
int sims = Math.Min(_simulations, 600);
int wins = 0;
for (int s = 0; s < sims; s++)
@ -52,7 +51,6 @@ namespace PdkFriendServer.Logic
return (double)wins / sims;
}
/// <summary>模拟一局只看胜负——不指定第一步出什么</summary>
private bool SimulateWin(TCardInfoPdkF[] myHand, PdkBotView view)
{
int myPos = view.MyPos;
@ -68,15 +66,14 @@ namespace PdkFriendServer.Logic
var unknown = allCards.Where(c => !myHand.Any(h => h.ID == c.ID)).ToList();
Shuffle(unknown);
int nextCnt = view.ShenYuCard[nextPos - 1];
int oppCnt = view.ShenYuCard[oppPos - 1];
hands[nextPos - 1].AddRange(unknown.Take(nextCnt));
hands[oppPos - 1].AddRange(unknown.Skip(nextCnt).Take(oppCnt));
hands[nextPos - 1].AddRange(unknown.Take(view.ShenYuCard[nextPos - 1]));
hands[oppPos - 1].AddRange(unknown.Skip(view.ShenYuCard[nextPos - 1])
.Take(view.ShenYuCard[oppPos - 1]));
// 模拟到有人出完
int current = myPos;
var maxPlay = new PlayOutCardPdkF(); // GameNum=0 → 新一轮
var maxPlay = new PlayOutCardPdkF();
int maxPlayer = 0;
int passStreak = 0;
for (int move = 0; move < 400; move++)
{
@ -89,7 +86,8 @@ namespace PdkFriendServer.Logic
if (tips != null && tips.Count > 0)
{
var pick = tips[0];
// P2修复: 优先多牌组合其次选最强最大GameNum
var pick = PickBestTip(tips, isNewRound);
maxPlay = new PlayOutCardPdkF
{
Pos = (byte)current,
@ -98,9 +96,35 @@ namespace PdkFriendServer.Logic
Ids = pick.Select(c => c.ID).ToArray()
};
maxPlayer = current;
passStreak = 0;
foreach (var id in maxPlay.Ids)
curHand.RemoveAll(c => c.ID == id);
}
else
{
passStreak++;
// 防止死循环: 三轮全 pass 则强制最大出牌
if (passStreak >= 2)
{
var forceTips = PdkCardAlgorithm.GetTipCard(
new PlayOutCardPdkF(), curHand.ToArray(), null, false);
if (forceTips != null && forceTips.Count > 0)
{
var pick = PickBestTip(forceTips, true);
maxPlay = new PlayOutCardPdkF
{
Pos = (byte)current,
GameNum = pick[0].GameNum,
Type = CardType1.DanZhang,
Ids = pick.Select(c => c.ID).ToArray()
};
maxPlayer = current;
foreach (var id in maxPlay.Ids)
curHand.RemoveAll(c => c.ID == id);
}
passStreak = 0;
}
}
current = current % 3 + 1;
}
@ -108,15 +132,15 @@ namespace PdkFriendServer.Logic
&& hands[myPos - 1].Count <= hands[oppPos - 1].Count;
}
// ---- 出牌决策 ----
public PlayOutCardPdkF DecidePlay(PdkBotView view)
{
var hand = view.MyHand;
if (hand.Length == 0)
return Pass(view);
if (hand.Length == 0) return Pass(view);
bool isFirstPlay = view.MaxPlayCard.GameNum <= 0;
// 枚举所有合法出牌
var tips = PdkCardAlgorithm.GetTipCard(
view.MaxPlayCard, hand, null, view.Rule.AAAIsZhaDan);
@ -131,7 +155,6 @@ namespace PdkFriendServer.Logic
if (candidates.Count == 1)
return candidates[0].play;
// 对每个候选做蒙特卡洛评估
int unknownCount = 48 - hand.Length - CountPlayed(hand, view);
int simsPerCandidate = Math.Max(50, _simulations / candidates.Count);
@ -170,38 +193,29 @@ namespace PdkFriendServer.Logic
int nextPos = myPos % 3 + 1;
int oppPos = (myPos + 1) % 3 + 1;
// 构建 3 人手牌:我已知,对手随机分配
var hands = new List<TCardInfoPdkF>[3];
hands[myPos - 1] = new List<TCardInfoPdkF>(myHand);
hands[nextPos - 1] = new List<TCardInfoPdkF>();
hands[oppPos - 1] = new List<TCardInfoPdkF>();
// 生成 48 张标准牌,排除我的手牌
var allCards = GenerateDeck();
var unknown = allCards.Where(c => !myHand.Any(h => h.ID == c.ID)).ToList();
Shuffle(unknown);
int nextCnt = view.ShenYuCard[nextPos - 1];
int oppCnt = view.ShenYuCard[oppPos - 1];
hands[nextPos - 1].AddRange(unknown.Take(nextCnt));
hands[oppPos - 1].AddRange(unknown.Skip(nextCnt).Take(oppCnt));
hands[nextPos - 1].AddRange(unknown.Take(view.ShenYuCard[nextPos - 1]));
hands[oppPos - 1].AddRange(unknown.Skip(view.ShenYuCard[nextPos - 1])
.Take(view.ShenYuCard[oppPos - 1]));
// 应用我的第一步出牌
var maxPlay = new PlayOutCardPdkF
{
Pos = (byte)myPos,
GameNum = myPlay.GameNum,
Type = myPlay.Type,
Ids = myPlay.Ids?.ToArray()
Pos = (byte)myPos, GameNum = myPlay.GameNum,
Type = myPlay.Type, Ids = myPlay.Ids?.ToArray()
};
if (myPlay.Type != CardType1.None && myPlay.Ids != null)
{
foreach (var id in myPlay.Ids)
hands[myPos - 1].RemoveAll(c => c.ID == id);
}
// 模拟到结束
int current = myPos % 3 + 1; // 下家先跟
int current = myPos % 3 + 1;
int maxPlayer = myPlay.Type == CardType1.None ? view.MaxPlayCard.Pos : myPos;
if (myPlay.Type == CardType1.None)
{
@ -210,60 +224,92 @@ namespace PdkFriendServer.Logic
}
int moveCount = 0;
int maxMoves = 500;
int passStreak = 0;
while (moveCount++ < maxMoves)
while (moveCount++ < 500)
{
var curHand = hands[current - 1];
if (curHand.Count == 0)
return current == myPos; // 我赢了
if (curHand.Count == 0) return current == myPos;
bool isNewRound = maxPlayer == current;
bool isNewRound = maxPlayer == current || maxPlay.GameNum == 0;
var curMax = isNewRound ? new PlayOutCardPdkF() : maxPlay;
var curTips = PdkCardAlgorithm.GetTipCard(curMax, curHand.ToArray(), null, false);
// 获取当前玩家的合法出牌
var curMax = isNewRound
? new PlayOutCardPdkF() // GameNum=0 means new round
: maxPlay;
var curTips = PdkCardAlgorithm.GetTipCard(
curMax, curHand.ToArray(), null, false);
if (isNewRound || (curTips != null && curTips.Count > 0))
if (curTips != null && curTips.Count > 0)
{
// 选第一个方案(最小牌)
var pick = isNewRound
? curHand.OrderBy(c => c.GameNum).First()
: curTips[0][0];
var ids = isNewRound
? new[] { pick.ID }
: curTips[0].Select(c => c.ID).ToArray();
var pick = PickBestTip(curTips, isNewRound);
maxPlay = new PlayOutCardPdkF
{
Pos = (byte)current,
GameNum = pick.GameNum,
Type = curTips?[0].Count == 1 ? CardType1.DanZhang : CardType1.DuiZi,
Ids = ids
GameNum = pick[0].GameNum,
Type = pick.Count == 1 ? CardType1.DanZhang : CardType1.DuiZi,
Ids = pick.Select(c => c.ID).ToArray()
};
maxPlayer = current;
foreach (var id in ids)
passStreak = 0;
foreach (var id in maxPlay.Ids)
curHand.RemoveAll(c => c.ID == id);
}
// else: pass, keep maxPlay unchanged
else
{
passStreak++;
if (passStreak >= 2)
{
var forceTips = PdkCardAlgorithm.GetTipCard(
new PlayOutCardPdkF(), curHand.ToArray(), null, false);
if (forceTips != null && forceTips.Count > 0)
{
var pick = PickBestTip(forceTips, true);
maxPlay = new PlayOutCardPdkF
{
Pos = (byte)current, GameNum = pick[0].GameNum,
Type = CardType1.DanZhang, Ids = pick.Select(c => c.ID).ToArray()
};
maxPlayer = current;
foreach (var id in maxPlay.Ids)
curHand.RemoveAll(c => c.ID == id);
}
passStreak = 0;
}
}
current = current % 3 + 1;
}
// 超时——比较剩余牌数
return hands[myPos - 1].Count <= hands[nextPos - 1].Count
&& hands[myPos - 1].Count <= hands[oppPos - 1].Count;
}
// ---- Rollout 策略 ----
/// <summary>
/// 从合法方案中选最优:优先出多张牌的组合,其次出最大的
/// P2修复: 之前只取 tips[0](最小单张),改为优先多牌组合
/// P0修复: 选最大(最后一项),模拟更强对手
/// </summary>
private static List<TCardInfoPdkF> PickBestTip(List<List<TCardInfoPdkF>> tips, bool isNewRound)
{
if (tips == null || tips.Count == 0) return null;
// 优先多牌组合2张>1张3张>2张等同数量选最大的
var best = tips[tips.Count - 1]; // 默认最大
int bestScore = best.Count * 10 + best[0].GameNum;
for (int i = tips.Count - 2; i >= 0; i--)
{
int score = tips[i].Count * 10 + tips[i][0].GameNum;
if (score > bestScore)
{
bestScore = score;
best = tips[i];
}
}
return best;
}
// ---- 工具 ----
private static int CountPlayed(TCardInfoPdkF[] myHand, PdkBotView view)
{
// 粗略估计已出牌:总牌 48 - 我的手牌 - 所有剩余牌
int total = 48 - myHand.Length;
foreach (var c in view.ShenYuCard) total -= c;
return Math.Max(0, total);
@ -273,17 +319,15 @@ namespace PdkFriendServer.Logic
{
var deck = new List<TCardInfoPdkF>();
int id = 1;
// 花色 1=黑桃 2=红桃 3=梅花 4=方块, 大小 1-13 (A=14, 2=16)
foreach (int flower in new[] { 1, 2, 3, 4 })
for (int num = 1; num <= 13; num++)
{
int gameNum = num switch
int gameNum = num switch { 1 => 14, 2 => 16, _ => num };
deck.Add(new TCardInfoPdkF
{
1 => 14, // A
2 => 16, // 2
_ => num
};
deck.Add(new TCardInfoPdkF { ID = (byte)id++, Flower = (byte)flower, GameNum = (byte)gameNum, GameState = 1 });
ID = (byte)id++, Flower = (byte)flower,
GameNum = (byte)gameNum, GameState = 1
});
}
return deck;
}
@ -301,10 +345,8 @@ namespace PdkFriendServer.Logic
{
return new PlayOutCardPdkF
{
Pos = src.Pos,
GameNum = src.GameNum,
Type = src.Type,
Ids = src.Ids?.ToArray()
Pos = src.Pos, GameNum = src.GameNum,
Type = src.Type, Ids = src.Ids?.ToArray()
};
}
@ -312,8 +354,7 @@ namespace PdkFriendServer.Logic
{
return new PlayOutCardPdkF
{
Pos = (byte)pos,
GameNum = cards[0].GameNum,
Pos = (byte)pos, GameNum = cards[0].GameNum,
Ids = cards.Select(x => x.ID).ToArray(),
Type = CardType1.DanZhang
};