📘

神经网络变得轻松（第四十三部分）：无需奖励函数精通技能·进阶篇

Q: 把多周期状态压成一维数组喂模型时，不同周期的顺序该怎么排才不会乱？

按周期从短到长或固定约定顺序拼接，并在注释里写死映射关系，避免训练与推理时错位。

Q: 持仓扫描得到的持仓量、盈亏这些数值，要归一化到什么范围再送强化学习状态？

建议线性缩放到 [-1,1] 或 [0,1]，并用账户净值做基准除权，防止绝对值尺度差异淹掉信号。

Q: 小布能不能帮我自动把账户状态机和判别器衔接的细节跑一遍并标出断点？

可以，小布能按你贴的网络结构扫描状态字段对应关系，高亮未衔接或维度不匹配的地方。

Q: 三网络反向传播时，判别器_loss 不降反而震荡，一般先调哪个参数？

先降判别器学习率或减小批量，确认标签噪声后再看特征层是否梯度消失。

Q: 账户状态机切换到达平仓态，但执行器没发单，调度器要查哪几个字段？

查状态机输出动作码、执行器可用保证金、以及调度器轮询间隔是否大于下单超时阈值。

第 2/3 篇

本章目录

判别器网络的层定义与参数落地
把多周期状态压进一维数组喂给模型
持仓扫描与强化学习状态喂入
把账户状态喂给调度器与执行器
账户状态机和判别器的衔接细节
三网络反向传播与训练进度输出

◍ 判别器网络的层定义与参数落地

这段构建逻辑给 GAN 里的判别器（discriminator）逐层铺结构，从输入到输出共 5 层，任何一层 CLayerDescription 实例化失败就直接 return false 并 delete 对象，避免悬空指针。输入层用 defNeuronBaseOCL，节点数 = HistoryBars * BarDescr + AccountDescr，window 设 0、激活函数 None，优化器统一走 ADAM。第一隐藏层是 defNeuronBatchNormOCL，节点数沿用上一层 prev_count，batch 写死 1000，做批量归一化但暂不开激活。第二、三层都是 defNeuronBaseOCL，各 256 节点，分别用 TANH 和 LReLU 激活；第四层 256 降到 NSkills 且 None 激活，第五层接 defNeuronSoftMaxOCL 输出 NSkills 类概率，step=1。在 MT5 里把 NSkills、HistoryBars 这类宏先定好，复制这段代码进 EA 的层初始化函数，编译跑通后能从日志看各层是否全部 Add 成功；外汇与贵金属杠杆品种波动剧烈，神经网络信号仅作概率参考，实盘前务必用历史数据验证。

MQL5 / C++

   class="kw">return class="kw">false;
    }
class=class="str">"cmt">//--- layer class="num">5
   if(!(descr = new CLayerDescription()))
      class="kw">return class="kw">false;
   descr.type = defNeuronSoftMaxOCL;
   descr.count = NSkills;
   descr.step = class="num">1;
   descr.optimization = ADAM;
   if(!scheduler.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return class="kw">false;
     } 
class=class="str">"cmt">//--- Discriminator
   discriminator.Clear();
class=class="str">"cmt">//--- Input layer
   if(!(descr = new CLayerDescription()))
      class="kw">return class="kw">false;
   descr.type = defNeuronBaseOCL;
   prev_count = descr.count = (HistoryBars * BarDescr + AccountDescr);
   descr.window = class="num">0;
   descr.activation = None;
   descr.optimization = ADAM;
   if(!discriminator.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return class="kw">false;
     }
class=class="str">"cmt">//--- layer class="num">1
   if(!(descr = new CLayerDescription()))
      class="kw">return class="kw">false;
   descr.type = defNeuronBatchNormOCL;
   descr.count = prev_count;
   descr.batch = class="num">1000;
   descr.activation = None;
   descr.optimization = ADAM;
   if(!discriminator.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return class="kw">false;
     }
class=class="str">"cmt">//--- layer class="num">2
   if(!(descr = new CLayerDescription()))
      class="kw">return class="kw">false;
   descr.type = defNeuronBaseOCL;
   descr.count = class="num">256;
   descr.optimization = ADAM;
   descr.activation = TANH;
   if(!discriminator.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return class="kw">false;
     }
class=class="str">"cmt">//--- layer class="num">3
   if(!(descr = new CLayerDescription()))
      class="kw">return class="kw">false;
   descr.type = defNeuronBaseOCL;
   descr.count = class="num">256;
   descr.optimization = ADAM;
   descr.activation = LReLU;
   if(!discriminator.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return class="kw">false;
     }
class=class="str">"cmt">//--- layer class="num">4
   if(!(descr = new CLayerDescription()))
      class="kw">return class="kw">false;
   descr.type = defNeuronBaseOCL;
   descr.count = NSkills;
   descr.optimization = ADAM;
   descr.activation = None;
   if(!discriminator.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return class="kw">false;
     }
class=class="str">"cmt">//--- layer class="num">5
   if(!(descr = new CLayerDescription()))
      class="kw">return class="kw">false;
   descr.type = defNeuronSoftMaxOCL;
   descr.count = NSkills;
   descr.step = class="num">1;
   descr.optimization = ADAM;
   if(!discriminator.Add(descr))
     {
      class="kw">delete descr;
      class="kw">return class="kw">false;
     }
class=class="str">"cmt">//---
   class="kw">return true;
   }

把多周期状态压进一维数组喂给模型

EA 在 OnTick 里先用 IsNewBar 拦掉旧 bar 的重复计算，只在新 K 线成型后跑一次特征组装，能少消耗至少 60% 的 tick 算力。下面的片段把 RSI、CCI、ATR、MACD 四个指标 refresh 后，循环 HistoryBars 根 K 线，每根用 12 个 float 槽位记录：收盘减开盘、最高减开盘、最低减开盘、tick_volume/1000、小时、星期、月份、以及四个指标主值/信号值。账户侧另开 5 个 float 存余额、净值、空闲保证金、保证金水平、浮动盈亏。这样 sState 就是一个定长向量，直接能丢进外部 Python 或 ONNX 模型做推理，不必再在 EA 内写复杂判断。注意外汇和贵金属杠杆高，这种特征工程只是把盘面结构化，不代表任何方向胜率，实盘前务必在 MT5 策略测试器用真实点差回测。

MQL5 / C++

class="type">void OnTick()
  {
class=class="str">"cmt">//---
   if(!IsNewBar())
      class="kw">return;
class=class="str">"cmt">//---
   class="type">int bars = CopyRates(Symb.Name(), TimeFrame, iTime(Symb.Name(), TimeFrame, class="num">1), HistoryBars, Rates);
   if(!ArraySetAsSeries(Rates, true))
      class="kw">return;
class=class="str">"cmt">//---
   RSI.Refresh();
   CCI.Refresh();
   ATR.Refresh();
   MACD.Refresh();
   class="type">MqlDateTime sTime;
   for(class="type">int b = class="num">0; b < (class="type">int)HistoryBars; b++)
     {
       class="type">class="kw">float open = (class="type">class="kw">float)Rates[b].open;
       TimeToStruct(Rates[b].time, sTime);
       class="type">class="kw">float rsi = (class="type">class="kw">float)RSI.Main(b);
       class="type">class="kw">float cci = (class="type">class="kw">float)CCI.Main(b);
       class="type">class="kw">float atr = (class="type">class="kw">float)ATR.Main(b);
       class="type">class="kw">float macd = (class="type">class="kw">float)MACD.Main(b);
       class="type">class="kw">float sign = (class="type">class="kw">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">//---
       sState.state[b * class="num">12] = (class="type">class="kw">float)Rates[b].close - open;
       sState.state[b * class="num">12 + class="num">1] = (class="type">class="kw">float)Rates[b].high - open;
       sState.state[b * class="num">12 + class="num">2] = (class="type">class="kw">float)Rates[b].low - open;
       sState.state[b * class="num">12 + class="num">3] = (class="type">class="kw">float)Rates[b].tick_volume / class="num">1000.0f;
       sState.state[b * class="num">12 + class="num">4] = (class="type">class="kw">float)sTime.hour;
       sState.state[b * class="num">12 + class="num">5] = (class="type">class="kw">float)sTime.day_of_week;
       sState.state[b * class="num">12 + class="num">6] = (class="type">class="kw">float)sTime.mon;
       sState.state[b * class="num">12 + class="num">7] = rsi;
       sState.state[b * class="num">12 + class="num">8] = cci;
       sState.state[b * class="num">12 + class="num">9] = atr;
       sState.state[b * class="num">12 + class="num">10] = macd;
       sState.state[b * class="num">12 + class="num">11] = sign;
     }
   sState.account[class="num">0] = (class="type">class="kw">float)AccountInfoDouble(ACCOUNT_BALANCE);
   sState.account[class="num">1] = (class="type">class="kw">float)AccountInfoDouble(ACCOUNT_EQUITY);
   sState.account[class="num">2] = (class="type">class="kw">float)AccountInfoDouble(ACCOUNT_MARGIN_FREE);
   sState.account[class="num">3] = (class="type">class="kw">float)AccountInfoDouble(ACCOUNT_MARGIN_LEVEL);
   sState.account[class="num">4] = (class="type">class="kw">float)AccountInfoDouble(ACCOUNT_PROFIT);
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;

「持仓扫描与强化学习状态喂入」

遍历账户所有持仓时，先用 PositionsTotal 拿到总数，再逐个比对 PositionGetSymbol(i) 与当前标的名称，非本品种直接 continue 跳过。这一层过滤很关键——多币种账户里混着 EURUSD 和 XAUUSD 时，不隔离品种会让多空手数统计彻底失真。进入 switch 后按 POSITION_TYPE 分流：BUY 分支累加 buy_value（手数）和 buy_profit（浮动盈亏），SELL 分支同理写入 sell_value / sell_profit。随后把这四项塞进 sState.account 的索引 5~8，供后续归一化使用。状态向量 State1 的拼装值得细看：账户维度的变化率（如 (account[0]-prev_balance)/prev_balance）直接 Add 进向量，而 one_hot 动作编码用 vector<float>::Zeros(NSkills) 生成后随机置 1，再 AddArray 拼到尾部。Actor.feedForward 拿到指针后做前向推理，getSample 抽出一个离散动作 act。 Train 函数里用 MathRand 做样本下标 tr 的均匀抽样，再用 MathRand()*MathRand()/32767^2 的平方分布抽时序位置 i，倾向把训练重心压在缓冲中段而非首尾。外汇与贵金属杠杆高，这套 RL 闭环在实盘前务必用 MT5 策略测试器跑通 Buffer 结构再上。

MQL5 / C++

class="type">int total = PositionsTotal();
for(class="type">int i = class="num">0; i < total; i++)
  {
    if(PositionGetSymbol(i) != Symb.Name())
      class="kw">continue;
    class="kw">switch((class="type">int)PositionGetInteger(POSITION_TYPE))
      {
       case POSITION_TYPE_BUY:
          buy_value += PositionGetDouble(POSITION_VOLUME);
          buy_profit += PositionGetDouble(POSITION_PROFIT);
          class="kw">break;
       case POSITION_TYPE_SELL:
          sell_value += PositionGetDouble(POSITION_VOLUME);
          sell_profit += PositionGetDouble(POSITION_PROFIT);
          class="kw">break;
      }
  }
sState.account[class="num">5] = (class="type">class="kw">float)buy_value;
sState.account[class="num">6] = (class="type">class="kw">float)sell_value;
sState.account[class="num">7] = (class="type">class="kw">float)buy_profit;
sState.account[class="num">8] = (class="type">class="kw">float)sell_profit;
State1.AssignArray(sState.state);
State1.Add((sState.account[class="num">0] - prev_balance) / prev_balance);
State1.Add(sState.account[class="num">1] / prev_balance);
State1.Add((sState.account[class="num">1] - prev_equity) / prev_equity);
State1.Add(sState.account[class="num">3] / class="num">100.0f);
State1.Add(sState.account[class="num">4] / prev_balance);
State1.Add(sState.account[class="num">5]);
State1.Add(sState.account[class="num">6]);
State1.Add(sState.account[class="num">7] / prev_balance);
State1.Add(sState.account[class="num">8] / prev_balance);
vector<class="type">class="kw">float> one_hot = vector<class="type">class="kw">float>::Zeros(NSkills);
class="type">int skill=(class="type">int)MathRound(MathRand()/class="num">32767.0*(NSkills-class="num">1));
one_hot[skill] = class="num">1;
State1.AddArray(one_hot);
if(!Actor.feedForward(GetPointer(State1), class="num">1, class="kw">false))
   class="kw">return;
class="type">int act = Actor.getSample();
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);
   class="type">uint ticks = GetTickCount();
   for(class="type">int iter = class="num">0; (iter < Iterations && !IsStopped()); iter ++)
     {
       class="type">int tr = (class="type">int)(((class="type">class="kw">double)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)) * (Buffer[tr].Total - class="num">2));
   State1.AssignArray(Buffer[tr].States[i].state);

◍ 把账户状态喂给调度器与执行器

这段逻辑干的事很直接：把上一根和当前根的账户余额、净值、持仓占用等字段拼成一维特征向量 State1，再先后丢进 Scheduler（调度网络）和 Actor（策略网络）做前向推理。注意 PrevBalance 用了 MathMax(i-1,0) 做边界保护，首根不会越界读负数索引。特征里 account[3]/100.0f 把百分比字段归一化到 0~1 区间，account[4]、[7]、[8] 都除以 PrevBalance 做权益占比缩放，这种处理能让神经网络对不同资金规模账户保持同一响应尺度。外汇与贵金属杠杆高，缩放不当可能让模型在回测里对小额账户过拟合。 Scheduler.getSample() 先选一个子策略编号，把 one-hot 结果 SchedulerResult 追加进 State1；随后 Actor.feedForward 输出 action（0 代表最小手数加仓）。两个网络推理前都用 IsStopped() 拦截终端关闭信号，失败则 PrintFormat 打函数名加行号并 break，方便在 MT5 Experts 日志里定位是哪一行前向传播挂了。 prof_1l 的计算取下一状态里 close-open 的归一化差值，再乘 SYMBOL_TRADE_TICK_VALUE_PROFIT 除以 SYMBOL_POINT，得到每标准点对应的浮动盈亏。你把这个片段直接贴进 EA 的训练循环，把 HistoryBars 调到 200 左右，就能在策略测试器里观察 State1 维度是否和你的网络输入层对齐。

MQL5 / C++

class="type">class="kw">float PrevBalance = Buffer[tr].States[MathMax(i - class="num">1, class="num">0)].account[class="num">0];
class="type">class="kw">float PrevEquity = Buffer[tr].States[MathMax(i - class="num">1, class="num">0)].account[class="num">1];
State1.Add((Buffer[tr].States[i].account[class="num">0] - PrevBalance) / PrevBalance);
State1.Add(Buffer[tr].States[i].account[class="num">1] / PrevBalance);
State1.Add((Buffer[tr].States[i].account[class="num">1] - PrevEquity) / PrevEquity);
State1.Add(Buffer[tr].States[i].account[class="num">3] / class="num">100.0f);
State1.Add(Buffer[tr].States[i].account[class="num">4] / PrevBalance);
State1.Add(Buffer[tr].States[i].account[class="num">5]);
State1.Add(Buffer[tr].States[i].account[class="num">6]);
State1.Add(Buffer[tr].States[i].account[class="num">7] / PrevBalance);
State1.Add(Buffer[tr].States[i].account[class="num">8] / PrevBalance);
if(IsStopped())
  {
   PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
   class="kw">break;
   }
if(!Scheduler.feedForward(GetPointer(State1), class="num">1, class="kw">false))
  {
   PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
   class="kw">break;
   }
class="type">int skill = Scheduler.getSample();
SchedulerResult = vector<class="type">class="kw">float>::Zeros(NSkills);
SchedulerResult[skill] = class="num">1;
State1.AddArray(SchedulerResult);
if(IsStopped())
  {
   PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
   class="kw">break;
   }
if(!Actor.feedForward(GetPointer(State1), class="num">1, class="kw">false))
  {
   PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
   class="kw">break;
   }
class="type">int action = Actor.getSample();
State1.AssignArray(Buffer[tr].States[i + class="num">1].state);
vector<class="type">class="kw">float> account;
account.Assign(Buffer[tr].States[i].account);
class="type">int bar = (HistoryBars - class="num">1) * BarDescr;
class="type">class="kw">double cl_op = Buffer[tr].States[i + class="num">1].state[bar];
class="type">class="kw">double prof_1l = SymbolInfoDouble(_Symbol, SYMBOL_TRADE_TICK_VALUE_PROFIT) * cl_op /
                 SymbolInfoDouble(_Symbol, SYMBOL_POINT);
class="kw">switch(action)
  {
   case class="num">0:
     account[class="num">5] += (class="type">class="kw">float)SymbolInfoDouble(_Symbol, SYMBOL_VOLUME_MIN);

账户状态机和判别器的衔接细节

上面这段 switch 结构在维护一个浮点数组 account[]，用下标区分余额、权益、浮动盈亏等字段。case 0 与 case 3 都按 prof_1l 对 account[5]（多仓手数）和 account[6]（空仓手数）折算盈亏，再写回 account[7]、account[8]，最后用 account[4]=account[7]+account[8] 汇总，account[1]=account[0]+account[4] 更新权益。 case 1 在开仓侧多了一句：account[6] += (float)SymbolInfoDouble(_Symbol, SYMBOL_VOLUME_MIN)，也就是空仓手数按当前品种最小交易量递增一档；外汇与贵金属品种的最小交易量常为 0.01 手，这一步直接受合约规格约束。 case 2 是清算分支：把 account[4] 并入 account[0]，重置 account[1]、account[2]，并用 for(bar=3; bar<AccountDescr; bar++) account[bar]=0 把后续状态槽清零，避免上一段历史污染下一轮。循环尾部把 PrevBalance/PrevEquity 取出后，用 (account[0]-PrevBalance)/PrevBalance 等 9 个比值塞进 State1 向量，再交给 Discriminator.feedForward 做前向推理；若返回 false 就 PrintFormat 打出函数名与行号并 break。开 MT5 把这段贴进 EA 的回测框架，改 SYMBOL_VOLUME_MIN 为 SYMBOL_VOLUME_STEP 可观察加仓粒度变化对 State1 分布的影响。

MQL5 / C++

account[class="num">7] += account[class="num">5] * (class="type">class="kw">float)prof_1l;
account[class="num">8] -= account[class="num">6] * (class="type">class="kw">float)prof_1l;
account[class="num">4] = account[class="num">7] + account[class="num">8];
account[class="num">1] = account[class="num">0] + account[class="num">4];
class="kw">break;
case class="num">1:
   account[class="num">6] += (class="type">class="kw">float)SymbolInfoDouble(_Symbol, SYMBOL_VOLUME_MIN);
   account[class="num">7] += account[class="num">5] * (class="type">class="kw">float)prof_1l;
   account[class="num">8] -= account[class="num">6] * (class="type">class="kw">float)prof_1l;
   account[class="num">4] = account[class="num">7] + account[class="num">8];
   account[class="num">1] = account[class="num">0] + account[class="num">4];
   class="kw">break;
case class="num">2:
   account[class="num">0] += account[class="num">4];
   account[class="num">1] = account[class="num">0];
   account[class="num">2] = account[class="num">0];
   for(bar = class="num">3; bar < AccountDescr; bar++)
      account[bar] = class="num">0;
   class="kw">break;
case class="num">3:
   account[class="num">7] += account[class="num">5] * (class="type">class="kw">float)prof_1l;
   account[class="num">8] -= account[class="num">6] * (class="type">class="kw">float)prof_1l;
   account[class="num">4] = account[class="num">7] + account[class="num">8];
   account[class="num">1] = account[class="num">0] + account[class="num">4];
   class="kw">break;
}
PrevBalance = Buffer[tr].States[i].account[class="num">0];
PrevEquity = Buffer[tr].States[i].account[class="num">1];
State1.Add((account[class="num">0] - PrevBalance) / PrevBalance);
State1.Add(account[class="num">1] / PrevBalance);
State1.Add((account[class="num">1] - PrevEquity) / PrevEquity);
State1.Add(account[class="num">3] / class="num">100.0f);
State1.Add(account[class="num">4] / PrevBalance);
State1.Add(account[class="num">5]);
State1.Add(account[class="num">6]);
State1.Add(account[class="num">7] / PrevBalance);
State1.Add(account[class="num">8] / PrevBalance);
class=class="str">"cmt">//---
if(!Discriminator.feedForward(GetPointer(State1), class="num">1, class="kw">false))
   {
    PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
    class="kw">break;
   }

「三网络反向传播与训练进度输出」

这段训练收口逻辑把 Actor、Discriminator、Scheduler 三个网络的误差回传串在了一起，每轮先取判别器与动作网络结果，再用分类交叉熵（LOSS_CCE）给 Actor 算损失。回传顺序很关键：Actor 先用 State1 做 backProp，随后 Discriminator 独立回传，最后 Scheduler 的损失被账户余额变化率加权——也就是 (account[0]-PrevBalance)/PrevBalance，让收益波动直接渗入调度网络梯度。为防止界面卡死，代码用 GetTickCount 做了 500 毫秒节流：超过这个间隔才用 Comment 打印一次各网络最近平均误差，格式精确到 8 位小数，训练百分比保留 2 位。收尾阶段清空 Comment，并把 Scheduler 与 Discriminator 的最终平均误差用 PrintFormat 打到日志，精度 10.7f，随后调用 ExpertRemove 让 EA 自行卸载，整个强化学习训练过程在 MT5 中便跑完一轮。

MQL5 / C++

   }
      Discriminator.getResults(DiscriminatorResult);
      Actor.getResults(ActorResult);
      ActorResult[action] = DiscriminatorResult.Loss(SchedulerResult, LOSS_CCE);
      Result.AssignArray(ActorResult);
      State1.AddArray(SchedulerResult);
      if(!Actor.backProp(Result, DiscountFactor, GetPointer(State1), class="num">1, class="kw">false))
        {
         PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
         class="kw">break;
        }
      Result.AssignArray(SchedulerResult);
      if(!Discriminator.backProp(Result))
        {
         PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
         class="kw">break;
        }
      Result.AssignArray(SchedulerResult * ((account[class="num">0] - PrevBalance) / PrevBalance));
      if(!Scheduler.backProp(Result, DiscountFactor, GetPointer(State1), class="num">1, class="kw">false))
        {
         PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
         class="kw">break;
        }
      if(GetTickCount() - ticks > class="num">500)
        {
         class="type">class="kw">string str = StringFormat("%-15s %class="num">5.2f%% -> Error %class="num">15.8f\n",
                                   "Scheduler", iter * class="num">100.0 / (class="type">class="kw">double)(Iterations), Scheduler.getRecentAverageError());
         str += StringFormat("%-15s %class="num">5.2f%% -> Error %class="num">15.8f\n",
                             "Discriminator",  iter * class="num">100.0 / (class="type">class="kw">double)(Iterations), Discriminator.getRecentAverageError());
         Comment(str);
         ticks = GetTickCount();
        }
     }
   Comment("");
class=class="str">"cmt">//---
   PrintFormat("%s -> %d -> %-15s %class="num">10.7f", __FUNCTION__, __LINE__, "Scheduler", Scheduler.getRecentAverageError());
   PrintFormat("%s -> %d -> %-15s %class="num">10.7f", __FUNCTION__, __LINE__, "Discriminator", Discriminator.getRecentAverageError());
   ExpertRemove();
class=class="str">"cmt">//---
   }