feat(PX): 自训练v1.5 P0准备 — NHN模式 + 总分法胜率解析 + 评估管线

CB-01: Program.cs 加 --nhn 模式 [NEURAL,ISMCTS,NEURAL] - 3人跑得快 N插花坐排列 - Rotations 数组补全为4种排列 (IIN/NII/INI/NHN) - --nhn 参数: 不生成CSV, 直接评估 CB-02: 胜率解析统一总分法 - TrainingCollector 加 ScoreAccumulator/BotWins/BotGames 静态累加器 - GameOver 无条件调用 AccumulateScore (eval + 训练模式均记录) - stress_summary.txt 输出 #BOT_SCORES: bot=wins/games,wr=XX%,score=YYY - orchestrator eval 解析改用 #BOT_SCORES 格式, 正则回退逻辑移除管线: orchestrator eval→stress_summary→总分法解析, 闭环可验证
2026-07-20 15:40:59 +08:00
parent e777aef687
commit 7566fe99a1
4 changed files with 85 additions and 15 deletions
--- a/PdkFriendServer/Logic/PdkGameMain.cs
+++ b/PdkFriendServer/Logic/PdkGameMain.cs
@ -492,7 +492,6 @@ namespace PdkFriendServer.Logic
                #endif

                            // 训练数据收集：输出本局所有 ISMCTS 决策
-                            if (TrainingCollector.Enabled)
                            {
                                float[] rewards = new float[GamePack.Info.PlayNum];
                                for (int i = 0; i < GamePack.Info.PlayNum; i++)
@ -501,6 +500,8 @@ namespace PdkFriendServer.Logic
                                    float s = GamePack.Info.Score[i];
                                    rewards[i] = s > 0 ? 1f : s < 0 ? -1f : 0f;
                                }
+                                var rawScores = GamePack.Info.Score.Select(s => (float)s).ToArray();
+                                TrainingCollector.AccumulateScore(BotTypes, rawScores);
                                TrainingCollector.FlushGame(_flushGameCounter++, rewards, GamePack.Info.OverPos, BotTypes);
                            }

--- a/PdkFriendServer/Logic/TrainingCollector.cs
+++ b/PdkFriendServer/Logic/TrainingCollector.cs
@ -16,6 +16,15 @@ namespace PdkFriendServer.Logic
        private static readonly List<TrainingRecord> _records = new List<TrainingRecord>();
        private static readonly string _sessionPrefix = DateTime.Now.ToString("yyyyMMdd_HHmm");

+        /// <summary>总分累加器：botType → 累积 Score（用于评估统计）</summary>
+        public static readonly Dictionary<string, float> ScoreAccumulator = new Dictionary<string, float>();
+
+        /// <summary>每 bot 的胜局计数：botType → wins</summary>
+        public static readonly Dictionary<string, int> BotWins = new Dictionary<string, int>();
+
+        /// <summary>每 bot 的总局数：botType → total games</summary>
+        public static readonly Dictionary<string, int> BotGames = new Dictionary<string, int>();
+
        private static readonly int[] Ranks = { 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16 };
        private const int NumRanks = 13;
        private const int NumSuits = 4;
@ -58,6 +67,42 @@ namespace PdkFriendServer.Logic
            _records.Clear();
        }

+        /// <summary>
+        /// 累加 bot 得分：botType → Score 总和 + win/game 计数（总分法评估用）
+        /// </summary>
+        public static void AccumulateScore(string[] botTypes, float[] rawScores)
+        {
+            if (botTypes == null || rawScores == null) return;
+            for (int i = 0; i < botTypes.Length && i < rawScores.Length; i++)
+            {
+                string t = botTypes[i];
+                if (!ScoreAccumulator.ContainsKey(t))
+                    ScoreAccumulator[t] = 0;
+                ScoreAccumulator[t] += rawScores[i];
+
+                if (!BotGames.ContainsKey(t))
+                    BotGames[t] = 0;
+                BotGames[t]++;
+
+                if (rawScores[i] > 0)
+                {
+                    if (!BotWins.ContainsKey(t))
+                        BotWins[t] = 0;
+                    BotWins[t]++;
+                }
+            }
+        }
+
+        /// <summary>
+        /// 重置总分累加器
+        /// </summary>
+        public static void ResetAccumulator()
+        {
+            ScoreAccumulator.Clear();
+            BotWins.Clear();
+            BotGames.Clear();
+        }
+
        /// <summary>
        /// 状态编码：与 Python encode_game_state 完全一致。
        /// </summary>
--- a/hjha-console/Program.cs
+++ b/hjha-console/Program.cs
@ -8,10 +8,11 @@ namespace hjha_console
    class Program
    {
        public static int RotateInterval = 0; // 0=disabled, N=rotate every N rounds
-        private static readonly string[][] Rotations = new[] {    // 3种排列
+        private static readonly string[][] Rotations = new[] {    // 4种排列，覆盖所有 NHN/NNH/HNN 组合
            new[] { "ISMCTS", "ISMCTS", "NEURAL" },
            new[] { "NEURAL", "ISMCTS", "ISMCTS" },
-            new[] { "ISMCTS", "NEURAL", "ISMCTS" }
+            new[] { "ISMCTS", "NEURAL", "ISMCTS" },
+            new[] { "NEURAL", "ISMCTS", "NEURAL" }    // NHN: N插花坐
        };

        static void Main(string[] args)
@ -22,6 +23,7 @@ namespace hjha_console
            if (args.Length > 0 && args[0] == "--mix-ismcts") { PdkGameMain.BotTypes = new[] { "ISMCTS", "ISMCTS", "ISMCTS" }; TrainingCollector.Enabled = true; args = new[] { "--stress", args.Length > 1 ? args[1] : "10" }; }
            if (args.Length > 0 && args[0] == "--bots") { PdkGameMain.BotTypes = args[1].Split(','); TrainingCollector.Enabled = true; args = args.Length > 2 ? new[] { "--stress", args[2] } : new[] { "--stress", "10" }; }
            if (args.Length > 0 && args[0] == "--n2h2v5") { PdkGameMain.BotTypes = new[] { "ISMCTS", "ISMCTS", "NEURAL_V5" }; TrainingCollector.Enabled = false; args = new[] { "--stress", args.Length > 1 ? args[1] : "10" }; }
+            if (args.Length > 0 && args[0] == "--nhn") { PdkGameMain.BotTypes = new[] { "NEURAL", "ISMCTS", "NEURAL" }; TrainingCollector.Enabled = false; args = new[] { "--stress", args.Length > 1 ? args[1] : "10" }; }
            if (args.Length > 0 && args[0] == "--stress")
            {
                int rounds = args.Length > 1 ? int.Parse(args[1]) : 100;
@ -50,6 +52,9 @@ namespace hjha_console
            File.Delete("stress_errors.txt");
            File.Delete("stress_summary.txt");

+            // 清空总分累加器
+            TrainingCollector.ResetAccumulator();
+
            Console.WriteLine($"=== PdkBot Stress Test: {totalRounds} rounds ===");
            Console.WriteLine($"Logs: {logDir}/");
            Console.WriteLine($"Start: {DateTime.Now:HH:mm:ss}");
@ -136,11 +141,22 @@ namespace hjha_console
            Console.WriteLine($"Rounds: {completed}/{totalRounds} completed, {errors} errors");
            Console.WriteLine($"Total time: {totalSw.Elapsed.TotalMinutes:F1} min");
            Console.WriteLine($"Avg per round: {totalSw.Elapsed.TotalSeconds / totalRounds:F1}s");
-            File.WriteAllText("stress_summary.txt",
-                $"Rounds: {completed}/{totalRounds}\n" +
-                $"Errors: {errors}\n" +
-                $"Total: {totalSw.Elapsed.TotalMinutes:F1}min\n" +
-                $"Avg: {totalSw.Elapsed.TotalSeconds / totalRounds:F1}s/round\n");
+            var sb = new System.Text.StringBuilder();
+            sb.AppendLine($"Rounds: {completed}/{totalRounds}");
+            sb.AppendLine($"Errors: {errors}");
+            sb.AppendLine($"Total: {totalSw.Elapsed.TotalMinutes:F1}min");
+            sb.AppendLine($"Avg: {totalSw.Elapsed.TotalSeconds / totalRounds:F1}s/round");
+            // Bot 得分统计（总分法：累积 Score 总和 + 胜局数）
+            sb.AppendLine("#BOT_SCORES");
+            foreach (var t in TrainingCollector.ScoreAccumulator.Keys)
+            {
+                int wins = TrainingCollector.BotWins.ContainsKey(t) ? TrainingCollector.BotWins[t] : 0;
+                int games = TrainingCollector.BotGames.ContainsKey(t) ? TrainingCollector.BotGames[t] : 0;
+                float score = TrainingCollector.ScoreAccumulator[t];
+                float wr = games > 0 ? (float)wins / games * 100f : 0f;
+                sb.AppendLine($"{t}={wins}/{games},wr={wr:F1}%,score={score:F0}");
+            }
+            File.WriteAllText("stress_summary.txt", sb.ToString());
        }
    }
 }
--- a/scripts/training_orchestrator.py
+++ b/scripts/training_orchestrator.py
@ -209,13 +209,21 @@ def train_v(prev_winrate=None):
            with open(sf) as f:
                summary = f.read()
            log(f"  {summary.strip()}")
-            # parse Neural win rate from summary
-            m = re.search(r'NEURAL.*?(\d+)\s*wins?.*?(\d+)\s*games?', summary, re.IGNORECASE)
-            if not m:
-                m = re.search(r'position.*?(\d+).*?wins?.*?(\d+)', summary, re.IGNORECASE)
-            if m:
-                wins, games = int(m.group(1)), int(m.group(2))
-                winrate = wins / games * 100 if games > 0 else 0.0
+            # parse Neural win rate from #BOT_SCORES (总分法: 每局 Score>0 算赢)
+            in_bot_scores = False
+            for line in summary.splitlines():
+                line = line.strip()
+                if line == "#BOT_SCORES":
+                    in_bot_scores = True
+                    continue
+                if in_bot_scores and line.startswith("NEURAL="):
+                    # Format: NEURAL=wins/games,wr=XX.X%,score=YYY
+                    m = re.search(r'wr=(\d+\.?\d*)%', line)
+                    if m:
+                        winrate = float(m.group(1))
+                    break
+            if winrate == 0.0:
+                log("  ⚠️ BOT_SCORES not found, fallback to legacy parse")
        log(f"  Neural winrate: {winrate:.1f}%")
    except Exception as e:
        log(f"  eval failed: {e}")