4 Commits

Author	SHA1	Message	Date
xiaoou	f66f56d0f8	feat: ISMCTS q-value as training label ... C#: TrainingCollector新增q_value列, IsmctsBot传入winRate CSV: reward→q_value, ISMCTS状态带真实评估值, Neural填0 Python: 优先取ISMCTS状态+评估值训练, 无q_value的Neural状态跳过 旧CSV格式(reward)自动fallback到Neural-only过滤 信号从'这局谁赢'改为'ISMCTS评估这步有多好'	2026-07-13 11:38:42 +08:00
xiaoou	22acc3b2ca	feat: CSV加player列+#RESULT汇总 — 训练中直接看胜率进化	2026-07-13 08:40:14 +08:00
xiaoou	e5e1f4818e	fix: V-learning 出牌率 0%→81% — MakePlay Type=None导致所有候选被当pass ... 根因: MakePlay(Type=CardType1.None) → play.Type==None 对所有候选为true → afterHand保持原样(未出牌) → 所有候选V值相同 → 始终选第一个(pass) 修复: - play.Type→play.Ids.Length判断是否真pass - EncodeAfterPlay更新ShenYuCard(出牌后剩余-1) - EncodeAfterPlay设MaxPlayCard为出的牌 - isPass用Ids==null而非Type 效果: Play 117 Pass 28(81%出牌) vs ISMCTS -108/10局	2026-07-13 03:03:36 +08:00
xiaoou	f43e47b0a6	feat: V-learning架构 — 改Q→V ... Q-learning问题: pass action Q值稳定偏高 → 永远pass V-learning修复: - 网络输出 14(Q-action)→1(V-state): P(win|state) - NeuralBot: 对每个候选编码出牌后state → 选V值最高的 - pass和出牌都是候选, 公平竞争(不是固定action slot) TrainingCollector: - 新增 EncodeAfterPlay(view, tracker, afterHand, playedIds) - EncodeState 支持 extraSeen 参数(模拟出牌时标记已见) Python: - train_v.py: V-learning训练器(load CSV→predict win/loss) - network.py: OUTPUT_DIM=14→1 验证: 200局0错误, V-model loss 0.81→0.0001 需更多数据覆盖状态空间	2026-07-13 02:52:48 +08:00