神经网络变得轻松（第三十三部分）：分布式 Q-学习中的分位数回归·综合运用

macd = (class="type">float)MACD.Main(bar_t);
sign = (class="type">float)MACD.Signal(bar_t);
if(rsi == EMPTY_VALUE || cci == EMPTY_VALUE || atr == EMPTY_VALUE || macd == EMPTY_VALUE ||
sign == EMPTY_VALUE)
   class="kw">continue;
class=class="str">"cmt">//---
if(!State2.Add((class="type">float)Rates[bar_t].close - open) || !State2.Add((class="type">float)Rates[bar_t].high - open) ||
!State2.Add((class="type">float)Rates[bar_t].low - open) || !State2.Add((class="type">float)Rates[bar_t].tick_volume / class="num">1000.0f) ||
   !State2.Add(sTime.hour) || !State2.Add(sTime.day_of_week) || !State2.Add(sTime.mon) ||
   !State2.Add(rsi) || !State2.Add(cci) || !State2.Add(atr) || !State2.Add(macd) || !State2.Add(sign))
   {
    PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
    class="kw">break;
   }
     }
   if(IsStopped())
     {
      PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
      ExpertRemove();
      class="kw">return;
     }
   if(State1.Total() < (class="type">int)HistoryBars * class="num">12 ||
      (use_target && State2.Total() < (class="type">int)HistoryBars * class="num">12))
      class="kw">continue;
   if(!StudyNet.feedForward(GetPointer(State1), class="num">12, true))
      class="kw">return;
   Rewards.BufferInit(Actions, class="num">0);
   class="type">class="kw">double reward = Rates[i].close - Rates[i].open;
   if(reward >= class="num">0)
     {
      if(!Rewards.Update(class="num">0, (class="type">float)(class="num">2 * reward)))
       class="kw">return;
      if(!Rewards.Update(class="num">1, (class="type">float)(-class="num">5 * reward)))
       class="kw">return;
      if(!Rewards.Update(class="num">2, (class="type">float) - reward))
       class="kw">return;
     }
   else
     {
      if(!Rewards.Update(class="num">0, (class="type">float)(class="num">5 * reward)))
       class="kw">return;

if(!Rewards.Update(class="num">1, (class="type">float)(-class="num">2 * reward)))
   class="kw">return;
if(!Rewards.Update(class="num">2, (class="type">float)reward))
   class="kw">return;
   }
if(!StudyNet.backProp(GetPointer(Rewards), DiscountFactor, (use_target ? GetPointer(State2) : NULL), class="num">12, true))
   class="kw">return;
if(GetTickCount() - ticks > class="num">500)
   {
   Comment(StringFormat("%.2f%%", batch * class="num">100.0 / (class="type">class="kw">double)(Batch * UpdateTarget)));
   ticks = GetTickCount();
   }
   }
if(StudyNet.getRecentAverageError() <= min_loss)
   {
   if(!StudyNet.UpdateTarget(FileName + ".nnw"))
      class="kw">continue;
   use_target = true;
   min_loss = StudyNet.getRecentAverageError();
   }
PrintFormat("Iteration %d, loss %.8f", iter, StudyNet.getRecentAverageError());
 }
 Comment("");
class=class="str">"cmt">//---
 PrintFormat("%s -> %d", __FUNCTION__, __LINE__);
 ExpertRemove();
}