🧠
神经网络变得简单（第 60 部分）：在线决策转换器（ODT）·进阶篇
（2/3）·经典决策转换器跑久了就亏，本文拆解 ODT 如何用在线微调补上这条裂缝
接上篇，我们继续深挖决策转换器的实战短板。离线训好的模型在测试段开头能盈利，越往后越容易连续无盈利交易，把前面利润吐光。定期重训太笨重，在线优调才是更顺手的解法。
EA 初始化时的指标与模型装载
这段初始化逻辑把 MACD、RSI、CCI、ATR 四个指标连同两组神经网络模型一次性挂到当前图表。MACD 默认参数沿用了经典组合：快线 12、慢线 26、信号线 9，价格源取收盘价，改这几个数就能直接换周期风格。 输入组里还有两个容易被忽略的开关：StudyIters=5 控制学习迭代次数，StudyPeriod=120 规定两次学习之间隔 120 根 K 线。做外汇或贵金属回测时，这两个值会明显改变样本外表现，属于高波动品种里必须手动压一压的变量。 OnInit 里先调 LoadTotalBase，再逐个 Create 指标句柄，任何一步失败就返回 INIT_FAILED，MT5 会直接在日志报哪一行断的。模型部分从 FileName 指向的 Act.nnw 与 RTG.nnw 读权重，若文件缺失，代码会新建空数组对象而不是崩溃——这点对调试很有用。
MQL5 / C++
input group "---- MACD ----" input class="type">int FastPeriod = class="num">12; class=class="str">"cmt">//Fast input class="type">int SlowPeriod = class="num">26; class=class="str">"cmt">//Slow input class="type">int SignalPeriod= class="num">9; class=class="str">"cmt">//Signal input ENUM_APPLIED_PRICE MACDPrice = PRICE_CLOSE; class=class="str">"cmt">//Applied price class=class="str">"cmt">//--- input class="type">int StudyIters = class="num">5; class=class="str">"cmt">//Iterations to Study input class="type">int StudyPeriod = class="num">120; class=class="str">"cmt">//Bars between Studies class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert initialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">int OnInit() { LoadTotalBase(); if(!Symb.Name(_Symbol)) class="kw">return INIT_FAILED; Symb.Refresh(); class=class="str">"cmt">//--- if(!RSI.Create(Symb.Name(), TimeFrame, RSIPeriod, RSIPrice)) class="kw">return INIT_FAILED; class=class="str">"cmt">//--- if(!CCI.Create(Symb.Name(), TimeFrame, CCIPeriod, CCIPrice)) class="kw">return INIT_FAILED; class=class="str">"cmt">//--- if(!ATR.Create(Symb.Name(), TimeFrame, ATRPeriod)) class="kw">return INIT_FAILED; class=class="str">"cmt">//--- if(!MACD.Create(Symb.Name(), TimeFrame, FastPeriod, SlowPeriod, SignalPeriod, MACDPrice)) class="kw">return INIT_FAILED; if(!RSI.BufferResize(NBarInPattern) || !CCI.BufferResize(NBarInPattern) || !ATR.BufferResize(NBarInPattern) || !MACD.BufferResize(NBarInPattern)) { PrintFormat("%s -> %d", __FUNCTION__, __LINE__); class="kw">return INIT_FAILED; } class=class="str">"cmt">//--- if(!Trade.SetTypeFillingBySymbol(Symb.Name())) class="kw">return INIT_FAILED; class=class="str">"cmt">//--- load models class="type">float temp; if(!Agent.Load(FileName + "Act.nnw", temp, temp, temp, dtStudied, true) || !RTG.Load(FileName + "RTG.nnw", dtStudied, true) || !AgentStudy.Load(FileName + "Act.nnw", temp, temp, temp, dtStudied, true) || !RTGStudy.Load(FileName + "RTG.nnw", dtStudied, true)) { PrintFormat("Can&class="macro">#x27;t load pretrained models"); CArrayObj *agent = new CArrayObj(); CArrayObj *rtg = new CArrayObj();
◍ 初始化失败点与模型落盘的细节
强化学习 EA 在 OnInit 里先建描述对象，若 CreateDescriptions 返回 false 会立刻 delete 掉 agent 与 rtg 并打印 "Can't create description of models"，直接 return INIT_FAILED。这一步卡住说明状态空间的描述生成有问题，常见原因是 NBarInPattern 或 BarDescr 配置和传入数据对不上。 紧接着的 Agent.Create / RTG.Create / AgentStudy.Create / RTGStudy.Create 任一失败也会走同一套清理逻辑并返回 INIT_FAILED。注意这里 delete 的是临时 agent、rtg 指针，真正参与推理的 study 对象在后面才用。 初始化尾段有两次维度校验：Result.Total() 必须等于 NActions，否则报 "The scope of the actor does not match the actions count"；GetLayerOutput(0, Result) 后 Total 必须等于 NRewards + BarDescr * NBarInPattern + AccountDescr + TimeDescription + NActions，差一个就 INIT_FAILED。这组加法就是你的状态向量长度，调参时改任一常量都要同步核对。 OnDeinit 里会把训练好的网络存成 Act.nnw 与 RTG.nnw，并调用 SaveTotalBase 落回经验池。回测中断或换品种前，确认这两个文件已生成，否则下次加载是冷启动。OnTick 开头只用 IsNewBar 过滤，非新 K 线直接 return，说明决策频率绑定图表周期，想提高响应就把 EA 挂到更低周期但注意外汇杠杆风险偏高。
MQL5 / C++
if(!CreateDescriptions(agent, rtg)) { class="kw">delete agent; class="kw">delete rtg; PrintFormat("Can&class="macro">#x27;t create description of models"); class="kw">return INIT_FAILED; } if(!Agent.Create(agent) || !RTG.Create(rtg) || !AgentStudy.Create(agent) || !RTGStudy.Create(rtg)) { class="kw">delete agent; class="kw">delete rtg; PrintFormat("Can&class="macro">#x27;t create models"); class="kw">return INIT_FAILED; } class="kw">delete agent; class="kw">delete rtg; class=class="str">"cmt">//--- } class=class="str">"cmt">//--- Agent.getResults(Result); if(Result.Total() != NActions) { PrintFormat("The scope of the actor does not match the actions count(%d <> %d)", NActions, Result.Total()); class="kw">return INIT_FAILED; } AgentResult = vector<class="type">float>::Zeros(NActions); class=class="str">"cmt">//--- Agent.GetLayerOutput(class="num">0, Result); if(Result.Total() != (NRewards + BarDescr * NBarInPattern + AccountDescr + TimeDescription + NActions)) { PrintFormat("Input size of Actor doesn&class="macro">#x27;t match state description(%d <> %d)", Result.Total(), (NRewards + BarDescr * NBarInPattern + AccountDescr + TimeDescription + NActions)); class="kw">return INIT_FAILED; } Agent.Clear(); RTG.Clear(); PrevBalance = AccountInfoDouble(ACCOUNT_BALANCE); PrevEquity = AccountInfoDouble(ACCOUNT_EQUITY); class=class="str">"cmt">//--- class="kw">return(INIT_SUCCEEDED); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert deinitialization function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnDeinit(const class="type">int reason) { class=class="str">"cmt">//--- AgentStudy.Save(FileName + "Act.nnw", class="num">0, class="num">0, class="num">0, TimeCurrent(), true); RTGStudy.Save(FileName + "RTG.nnw", TimeCurrent(), true); class="kw">delete Result; class="type">int total = ArraySize(Buffer); printf("Saving %d", MathMin(total + class="num">1, MaxReplayBuffer)); SaveTotalBase(); Print("Saved"); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Expert tick function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void OnTick() { class=class="str">"cmt">//--- if(!IsNewBar()) class="kw">return; class=class="str">"cmt">//---
「把多指标状态压进一维特征数组」
这段逻辑做的是把一段 K 线窗口里的价格行为与指标读数，统一塞进一个定长浮点数组，方便后续直接喂给模型或比对。先通过 CopyRates 拉取从倒数第 1 根 bar 起、共 NBarInPattern 根的历史数据，并强制把 Rates 设为序列排列（下标 0 为最新），若设置失败直接 return，避免后面错位。 随后逐个 Refresh 指标句柄（RSI/CCI/ATR/MACD/符号），保证读到的是当前帧数据。循环里每根 bar 占 BarDescr 个槽位，依次写入 close-open、high-open、low-open、tick_volume/1000、以及 rsi/cci/atr/macd/signal 九个值；任一指标返回 EMPTY_VALUE 就跳过该 bar，不污染特征。 账户层只抓了两个量：balance 与 equity，写进 account[0]、account[1]。持仓统计部分则遍历 PositionsTotal，按品种过滤后把多单 volume 累加进 buy_value、利润进 buy_profit，空单同理，multiplier 用 1/(60*60*10) 做时间衰减系数，可用来折算持仓时长权重。外汇与贵金属杠杆高，这类特征若用于实盘信号，须先在策略测试器跑历史样本验证分布稳定性。 复制下面代码到 MT5 的 EA 头里，把 NBarInPattern 调到你能覆盖的波段长度（例如 30），即可在观察器里看 sState.state 每根 bar 的九维切片。
MQL5 / C++
class="type">int bars = CopyRates(Symb.Name(), TimeFrame, iTime(Symb.Name(), TimeFrame, class="num">1), NBarInPattern, Rates); if(!ArraySetAsSeries(Rates, true)) class="kw">return; class=class="str">"cmt">//--- RSI.Refresh(); CCI.Refresh(); ATR.Refresh(); MACD.Refresh(); Symb.Refresh(); Symb.RefreshRates(); class=class="str">"cmt">//--- History data class="type">float atr = class="num">0; for(class="type">int b = class="num">0; b < (class="type">int)NBarInPattern; b++) { class="type">float open = (class="type">float)Rates[b].open; class="type">float rsi = (class="type">float)RSI.Main(b); class="type">float cci = (class="type">float)CCI.Main(b); atr = (class="type">float)ATR.Main(b); class="type">float macd = (class="type">float)MACD.Main(b); class="type">float sign = (class="type">float)MACD.Signal(b); if(rsi == EMPTY_VALUE || cci == EMPTY_VALUE || atr == EMPTY_VALUE || macd == EMPTY_VALUE || sign == EMPTY_VALUE) class="kw">continue; class=class="str">"cmt">//--- class="type">int shift = b * BarDescr; sState.state[shift] = (class="type">float)(Rates[b].close - open); sState.state[shift + class="num">1] = (class="type">float)(Rates[b].high - open); sState.state[shift + class="num">2] = (class="type">float)(Rates[b].low - open); sState.state[shift + class="num">3] = (class="type">float)(Rates[b].tick_volume / class="num">1000.0f); sState.state[shift + class="num">4] = rsi; sState.state[shift + class="num">5] = cci; sState.state[shift + class="num">6] = atr; sState.state[shift + class="num">7] = macd; sState.state[shift + class="num">8] = sign; } bState.AssignArray(sState.state); class=class="str">"cmt">//--- Account description sState.account[class="num">0] = (class="type">float)AccountInfoDouble(ACCOUNT_BALANCE); sState.account[class="num">1] = (class="type">float)AccountInfoDouble(ACCOUNT_EQUITY); class=class="str">"cmt">//--- class="type">class="kw">double buy_value = class="num">0, sell_value = class="num">0, buy_profit = class="num">0, sell_profit = class="num">0; class="type">class="kw">double position_discount = class="num">0; class="type">class="kw">double multiplyer = class="num">1.0 / (class="num">60.0 * class="num">60.0 * class="num">10.0); class="type">int total = PositionsTotal(); class="type">class="kw">datetime current = TimeCurrent(); for(class="type">int i = class="num">0; i < total; i++) { if(PositionGetSymbol(i) != Symb.Name()) class="kw">continue; class="type">class="kw">double profit = PositionGetDouble(POSITION_PROFIT); class="kw">switch((class="type">int)PositionGetInteger(POSITION_TYPE)) { case POSITION_TYPE_BUY: buy_value += PositionGetDouble(POSITION_VOLUME); buy_profit += profit; break;
把持仓状态压进神经网络的输入向量
这段逻辑干的事很直接：先按 POSITION_TYPE_SELL 分支把空单手数和浮动盈亏分别累加进 sell_value、sell_profit，再算一笔 position_discount——用持仓利润减去「当前时间减开仓时间」乘系数乘利润绝对值，等于给长周期持仓打了个时间折扣。 sState.account 数组下标 2~7 依次落了买量、卖量、买利、卖利、折扣、当前 K 线时间；随后 bState 把账户回撤率（(余额-前余额)/前余额）、权益占比、权益回撤、买卖量原值、买卖利占比、折扣占比共 8 个浮点塞进去，构成 RL 状态基底。 时间标签那段把 Rates[0].time 分别除以 2023 全年秒数、月线/周线/日线周期秒数，再跑 sin/cos(2πx)，等于手工编码了年/月/周/日的周期相位，喂给网络当季节性特征。 真正下注前，代码用 RTG.feedForward 和 Agent.feedForward 两层前向传播拿结果，更新 PrevBalance/PrevEquity，再从 Agent 取 temp 向量。temp[0] 与 temp[3] 做互斥裁剪（谁大减谁、小的归零）后，才用 LotsMin、LotsStep、StopsLevel*Point 约束下单手数和止损距——外汇贵金属杠杆高，这种裁剪能压住同向重复开仓的敞口膨胀。
MQL5 / C++
case POSITION_TYPE_SELL: sell_value += PositionGetDouble(POSITION_VOLUME); sell_profit += profit; break; } position_discount += profit - (current - PositionGetInteger(POSITION_TIME)) * multiplyer * MathAbs(profit); } sState.account[class="num">2] = (class="type">float)buy_value; sState.account[class="num">3] = (class="type">float)sell_value; sState.account[class="num">4] = (class="type">float)buy_profit; sState.account[class="num">5] = (class="type">float)sell_profit; sState.account[class="num">6] = (class="type">float)position_discount; sState.account[class="num">7] = (class="type">float)Rates[class="num">0].time; class=class="str">"cmt">//--- bState.Add((class="type">float)((sState.account[class="num">0] - PrevBalance) / PrevBalance)); bState.Add((class="type">float)(sState.account[class="num">1] / PrevBalance)); bState.Add((class="type">float)((sState.account[class="num">1] - PrevEquity) / PrevEquity)); bState.Add(sState.account[class="num">2]); bState.Add(sState.account[class="num">3]); bState.Add((class="type">float)(sState.account[class="num">4] / PrevBalance)); bState.Add((class="type">float)(sState.account[class="num">5] / PrevBalance)); bState.Add((class="type">float)(sState.account[class="num">6] / PrevBalance)); class=class="str">"cmt">//--- Time label class="type">class="kw">double x = (class="type">class="kw">double)Rates[class="num">0].time / (class="type">class="kw">double)(D&class="macro">#x27;class="num">2024.01.class="num">01&class="macro">#x27; - D&class="macro">#x27;class="num">2023.01.class="num">01&class="macro">#x27;); bState.Add((class="type">float)MathSin(class="num">2.0 * M_PI * x)); x = (class="type">class="kw">double)Rates[class="num">0].time / (class="type">class="kw">double)PeriodSeconds(PERIOD_MN1); bState.Add((class="type">float)MathCos(class="num">2.0 * M_PI * x)); x = (class="type">class="kw">double)Rates[class="num">0].time / (class="type">class="kw">double)PeriodSeconds(PERIOD_W1); bState.Add((class="type">float)MathSin(class="num">2.0 * M_PI * x)); x = (class="type">class="kw">double)Rates[class="num">0].time / (class="type">class="kw">double)PeriodSeconds(PERIOD_D1); bState.Add((class="type">float)MathSin(class="num">2.0 * M_PI * x)); class=class="str">"cmt">//--- Prev action bState.AddArray(AgentResult); class=class="str">"cmt">//--- Return to go if(!RTG.feedForward(GetPointer(bState))) class="kw">return; RTG.getResults(Result); bState.AddArray(Result); class=class="str">"cmt">//--- if(!Agent.feedForward(GetPointer(bState), class="num">1, false, (CBufferFloat*)NULL)) class="kw">return; class=class="str">"cmt">//--- PrevBalance = sState.account[class="num">0]; PrevEquity = sState.account[class="num">1]; class=class="str">"cmt">//--- vector<class="type">float> temp; Agent.getResults(temp); class=class="str">"cmt">//--- class="type">class="kw">double min_lot = Symb.LotsMin(); class="type">class="kw">double step_lot = Symb.LotsStep(); class="type">class="kw">double stops = MathMax(Symb.StopsLevel(), class="num">1) * Symb.Point(); if(temp[class="num">0] >= temp[class="num">3]) { temp[class="num">0] -= temp[class="num">3]; temp[class="num">3] = class="num">0; } else {
◍ 多空双向的仓位再平衡逻辑
这段控制块把 agent 算出的目标权重 temp[] 直接映射到实盘下单动作，买和卖各走一套对称判断。先看买侧：当 temp[0] 小于最小手数，或止盈/止损距离换算后小于等于 stops 阈值时，若有买仓就整仓平掉；否则按 step_lot 步进把目标手数对齐，并依 MaxTP、MaxSL 和点值算出 tp/sl 价格。 若已有买仓，先调用 TrailPosition 把止损止盈 trailing 到新位置；若当前买仓手数 buy_value 与目标 buy_lot 不一致，多出部分用 ClosePartial 减仓，不足部分用 Trade.Buy 补单。卖侧逻辑完全镜像，只是价格基于 Bid 且 sl 在 Bid 之上、tp 在 Bid 之下。 最后一段把状态写回强化学习结构：用 BarDescr*(NBarInPattern-1) 定位上一pattern 的末端 bar，把 bState 偏移量塞进 sState.rewards，并在无持仓时按 atr/PrevBalance 扣减奖励、有持仓则奖励清零。外汇与贵金属杠杆高，这类按权重自动加减仓的逻辑在滑点扩大时可能触发频繁往返交易，建议先在 MT5 策略测试器用点差 20 点以上的历史段跑一遍。
MQL5 / C++
temp[class="num">3] -= temp[class="num">0]; temp[class="num">0] = class="num">0; } class="type">float delta = MathAbs(AgentResult - temp).Sum(); AgentResult = temp; class=class="str">"cmt">//--- buy control if(temp[class="num">0] < min_lot || (temp[class="num">1] * MaxTP * Symb.Point()) <= stops || (temp[class="num">2] * MaxSL * Symb.Point()) <= stops) { if(buy_value > class="num">0) CloseByDirection(POSITION_TYPE_BUY); } else { class="type">class="kw">double buy_lot = min_lot + MathRound((class="type">class="kw">double)(temp[class="num">0] - min_lot) / step_lot) * step_lot; class="type">class="kw">double buy_tp = Symb.NormalizePrice(Symb.Ask() + temp[class="num">1] * MaxTP * Symb.Point()); class="type">class="kw">double buy_sl = Symb.NormalizePrice(Symb.Ask() - temp[class="num">2] * MaxSL * Symb.Point()); if(buy_value > class="num">0) TrailPosition(POSITION_TYPE_BUY, buy_sl, buy_tp); if(buy_value != buy_lot) { if(buy_value > buy_lot) ClosePartial(POSITION_TYPE_BUY, buy_value - buy_lot); else Trade.Buy(buy_lot - buy_value, Symb.Name(), Symb.Ask(), buy_sl, buy_tp); } } class=class="str">"cmt">//--- sell control if(temp[class="num">3] < min_lot || (temp[class="num">4] * MaxTP * Symb.Point()) <= stops || (temp[class="num">5] * MaxSL * Symb.Point()) <= stops) { if(sell_value > class="num">0) CloseByDirection(POSITION_TYPE_SELL); } else { class="type">class="kw">double sell_lot = min_lot + MathRound((class="type">class="kw">double)(temp[class="num">3] - min_lot) / step_lot) * step_lot;; class="type">class="kw">double sell_tp = Symb.NormalizePrice(Symb.Bid() - temp[class="num">4] * MaxTP * Symb.Point()); class="type">class="kw">double sell_sl = Symb.NormalizePrice(Symb.Bid() + temp[class="num">5] * MaxSL * Symb.Point()); if(sell_value > class="num">0) TrailPosition(POSITION_TYPE_SELL, sell_sl, sell_tp); if(sell_value != sell_lot) { if(sell_value > sell_lot) ClosePartial(POSITION_TYPE_SELL, sell_value - sell_lot); else Trade.Sell(sell_lot - sell_value, Symb.Name(), Symb.Bid(), sell_sl, sell_tp); } } class=class="str">"cmt">//--- class="type">int shift = BarDescr * (NBarInPattern - class="num">1); sState.rewards[class="num">0] = bState[shift]; sState.rewards[class="num">1] = bState[shift + class="num">1] - class="num">1.0f; if((buy_value + sell_value) == class="num">0) sState.rewards[class="num">2] -= (class="type">float)(atr / PrevBalance); else sState.rewards[class="num">2] = class="num">0; for(class="type">ulong i = class="num">0; i < NActions; i++) sState.action[i] = AgentResult[i];
「训练循环里的样本剔除与随机抽帧」
上面这段落在 OnTick 末尾调用 Train()，但真正干活的是 Train(void) 里的内存整理与迭代采样。先看头部：当 Base.Total 超过 StudyPeriod 才把当前 Base 压入 Buffer 并做累计收益处理，否则不进样本池。
样本清理有一段容易被忽略：clear = Base.Total + StudyPeriod - Buffer_Size，若 clear > 0 就调 Base.ClearFirstN(clear) 丢掉最早的状态，避免数组无限膨胀。MT5 里若 Buffer_Size 设太小，老样本被清太快，回测可能偏向近期行情。
随后用 count 变量压缩 Buffer：凡是 Buffer[i].Total < StudyPeriod 的残样直接前移覆盖，最后 ArrayResize 到 total_tr - count。这一步保证后续训练只跑满周期的样本。
迭代部分用 MathRand()/32767.0 先随机抽一条轨迹 tr，再用 MathRand()*MathRand()/32767^2 的偏置分布选起点 i，倾向落在轨迹中段而非首尾。若 i<0 就 iter-- 重抽，不浪费迭代次数。外汇与贵金属波动跳变多，这种中段采样可能降低极端边界样本干扰，但仍是概率性改善，实盘前请在 MT5 策略测试器用真实点差验证。
MQL5 / C++
if(!Base.Add(sState)) ExpertRemove(); class=class="str">"cmt">//--- if((Bars(_Symbol, TimeFrame) % StudyPeriod) == class="num">0) Train(); } class=class="str">"cmt">//+------------------------------------------------------------------+ class=class="str">"cmt">//| Train function | class=class="str">"cmt">//+------------------------------------------------------------------+ class="type">void Train(class="type">void) { class="type">int total_tr = ArraySize(Buffer); if(Base.Total >= StudyPeriod) if(ArrayResize(Buffer, total_tr + class="num">1) == (total_tr + class="num">1)) { Buffer[total_tr] = Base; Buffer[total_tr].CumRevards(); total_tr++; } class="type">int clear = Base.Total + StudyPeriod - Buffer_Size; if(clear > class="num">0) Base.ClearFirstN(clear); class=class="str">"cmt">//--- class="type">int count = class="num">0; for(class="type">int i = class="num">0; i < (total_tr + count); i++) { if(Buffer[i + count].Total < StudyPeriod) { count++; i--; class="kw">continue; } if(count > class="num">0) Buffer[i] = Buffer[i + count]; } if(count > class="num">0) { ArrayResize(Buffer, total_tr - count); total_tr = ArraySize(Buffer); } class="type">uint ticks = GetTickCount(); class=class="str">"cmt">//--- class="type">bool StopFlag = false; for(class="type">int iter = class="num">0; (iter < StudyIters && !IsStopped() && !StopFlag); iter ++) { class="type">int tr = (class="type">int)((MathRand() / class="num">32767.0) * (total_tr - class="num">1)); class="type">int i = (class="type">int)((MathRand() * MathRand() / MathPow(class="num">32767, class="num">2)) * MathMax(Buffer[tr].Total - class="num">2 * HistoryBars, MathMin(Buffer[tr].Total, class="num">20))); if(i < class="num">0) { iter--; class="kw">continue; } vector<class="type">float> Actions = vector<class="type">float>::Zeros(NActions); AgentStudy.Clear(); RTGStudy.Clear(); for(class="type">int state = i; state < MathMin(Buffer[tr].Total - class="num">2, class="type">int(i + HistoryBars * class="num">1.5)); state++) { class=class="str">"cmt">//--- History data bState.AssignArray(Buffer[tr].States[state].state); class=class="str">"cmt">//--- Account description
交给小布盯盘看回撤拐点
这些诊断小布盯盘的 AIGC 已内置，打开对应品种页即可看到模型收益曲线转弱的那一段，你再决定是否手动介入优调。
常见问题
经典 DT 只用离线轨迹训静态策略，ODT 在离线初训后接一段在线优调，用环境交互数据补状态和动作空间的覆盖盲区，策略倾向随盘面持续修正。
离线集回报偏低且只覆盖有限状态动作子空间，模型遇到分布外行情时预测动作可能失效，从而产生连续无盈利业务并回吐利润。
可以。文中提到操作期上下文长度可能短于训练期，K 作为超参数控制转换器可见的历史令牌步数，实盘可按延迟和资源权衡截断。
目前小布内置的是诊断与看盘辅助，ODT 的训练代码需你在 MT5 端自行部署；小布可帮你盯住策略回撤拐点，减少人工轮巡成本。
存在这种概率。外汇贵金属属高风险品种，在线步长与学习率需保守设，避免模型把短期噪声当成规律，实盘前建议用隔离样本回测。