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
156 lines
5.7 KiB
C#
156 lines
5.7 KiB
C#
using System;
|
||
using System.Collections.Generic;
|
||
using System.Linq;
|
||
using GameMessage;
|
||
using GameMessage.PaoDeKuaiF;
|
||
|
||
namespace PdkFriendServer.Logic
|
||
{
|
||
/// <summary>
|
||
/// 训练数据收集器 — 记录 (state, action, result) 三元组供 Python 训练。
|
||
/// 状态编码与 Python trainer 完全一致:13rank×4suit×7ch=364 float。
|
||
/// </summary>
|
||
public static class TrainingCollector
|
||
{
|
||
public static bool Enabled = false;
|
||
private static readonly List<TrainingRecord> _records = new List<TrainingRecord>();
|
||
|
||
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;
|
||
private const int NumChannels = 7;
|
||
public const int InputDim = NumChannels * NumRanks * NumSuits; // 364
|
||
|
||
private struct TrainingRecord
|
||
{
|
||
public float[] State;
|
||
public int Action; // 0-12 = rank index, 13 = pass
|
||
public int Player;
|
||
}
|
||
|
||
public static void Record(PdkBotView view, CardTracker tracker, int actionIdx, int player)
|
||
{
|
||
if (!Enabled) return;
|
||
var state = EncodeState(view, tracker);
|
||
_records.Add(new TrainingRecord { State = state, Action = actionIdx, Player = player });
|
||
}
|
||
|
||
/// <summary>
|
||
/// 游戏结束,输出本轮训练数据到文件。
|
||
/// rewards: 每个玩家的奖励 (1=赢, 0=第二, -1=第三)
|
||
/// </summary>
|
||
public static void FlushGame(int gameId, float[] rewards)
|
||
{
|
||
if (!Enabled || _records.Count == 0) return;
|
||
var path = $"training_data/game_{gameId:D5}.csv";
|
||
System.IO.Directory.CreateDirectory("training_data");
|
||
using var w = new System.IO.StreamWriter(path);
|
||
w.WriteLine("state,action,reward");
|
||
foreach (var r in _records)
|
||
{
|
||
float reward = r.Player >= 0 && r.Player < rewards.Length ? rewards[r.Player] : 0;
|
||
var stateStr = string.Join(",", r.State.Select(f => f.ToString("F6")));
|
||
w.WriteLine($"\"{stateStr}\",{r.Action},{reward:F3}");
|
||
}
|
||
_records.Clear();
|
||
}
|
||
|
||
/// <summary>
|
||
/// 状态编码:与 Python encode_game_state 完全一致。
|
||
/// </summary>
|
||
/// <summary>
|
||
/// 编码"出牌后"状态:把我刚出的牌也标记为已见
|
||
/// </summary>
|
||
public static float[] EncodeAfterPlay(PdkBotView view, CardTracker tracker,
|
||
TCardInfoPdkF[] afterHand, int[] playedIds)
|
||
{
|
||
var afterView = new PdkBotView
|
||
{
|
||
MyHand = afterHand,
|
||
MyPos = view.MyPos,
|
||
ShenYuCard = view.ShenYuCard,
|
||
MaxPlayCard = view.MaxPlayCard,
|
||
Rule = view.Rule,
|
||
};
|
||
return EncodeState(afterView, tracker, playedIds);
|
||
}
|
||
|
||
public static float[] EncodeState(PdkBotView view, CardTracker tracker, int[] extraSeen = null)
|
||
{
|
||
var state = new float[InputDim];
|
||
|
||
// ch0: my hand
|
||
foreach (var c in view.MyHand)
|
||
{
|
||
if (c.GameState != 1) continue;
|
||
int ri = RankToIdx(c.GameNum);
|
||
int si = SuitToIdx(c.ID);
|
||
if (ri >= 0 && si >= 0)
|
||
state[ri * NumSuits + si] = 1.0f;
|
||
}
|
||
|
||
// ch1: seen cards (from tracker) + extraSeen
|
||
var pool = tracker.GetUnknownPool(view.MyHand);
|
||
var unseenIds = new HashSet<int>(pool.Select(c => (int)c.ID));
|
||
if (extraSeen != null)
|
||
foreach (var id in extraSeen) unseenIds.Remove(id);
|
||
for (int suit = 1; suit <= 4; suit++)
|
||
{
|
||
for (int n = 1; n <= 13; n++)
|
||
{
|
||
int id = (suit - 1) * 13 + n;
|
||
if (!unseenIds.Contains(id))
|
||
{
|
||
int g = n == 1 ? 14 : n == 2 ? 16 : n;
|
||
int ri = RankToIdx(g);
|
||
if (ri >= 0)
|
||
state[ChOffset(1) + ri * NumSuits + (suit - 1)] = 1.0f;
|
||
}
|
||
}
|
||
}
|
||
|
||
// ch2-ch6: scalar info
|
||
int myCount = view.MyHand.Count(c => c.GameState == 1);
|
||
state[ChOffset(2)] = (float)myCount / 16.0f;
|
||
|
||
int opp1 = view.MyPos % 3;
|
||
state[ChOffset(3)] = (float)view.ShenYuCard[opp1] / 16.0f;
|
||
|
||
int opp2 = (view.MyPos + 1) % 3;
|
||
state[ChOffset(4)] = (float)view.ShenYuCard[opp2] / 16.0f;
|
||
|
||
state[ChOffset(5) + view.MyPos % NumRanks] = 1.0f;
|
||
|
||
if (view.MaxPlayCard.GameNum > 0 && view.MaxPlayCard.Pos != view.MyPos)
|
||
{
|
||
int lr = RankToIdx(view.MaxPlayCard.GameNum);
|
||
if (lr >= 0)
|
||
state[ChOffset(6) + lr] = 1.0f;
|
||
}
|
||
|
||
return state;
|
||
}
|
||
|
||
public static int RankToIdx(int gameNum) => gameNum switch
|
||
{
|
||
3 => 0, 4 => 1, 5 => 2, 6 => 3, 7 => 4, 8 => 5, 9 => 6,
|
||
10 => 7, 11 => 8, 12 => 9, 13 => 10, 14 => 11, 16 => 12,
|
||
_ => -1
|
||
};
|
||
|
||
public static int ActionToIdx(PlayOutCardPdkF play)
|
||
{
|
||
if (play.Type == CardType1.None || play.Ids == null || play.Ids.Length == 0)
|
||
return 13; // pass
|
||
return RankToIdx(play.GameNum);
|
||
}
|
||
|
||
private static int SuitToIdx(int id)
|
||
{
|
||
int suit = (id - 1) / 13;
|
||
return suit >= 0 && suit < 4 ? suit : -1;
|
||
}
|
||
|
||
private static int ChOffset(int ch) => ch * NumRanks * NumSuits;
|
||
}
|
||
} |