Машинное обучение в трейдинге: теория, модели, практика и алготорговля - страница 1067

Новый комментарий
 
Grigoriy Chaunin:
Пишите помедленней пожалуйста, я записываю. ) Может пора перейти на русский? А то мне неучу ничего не понятно.
про объемы согласен
 
Maxim Dmitrievsky :

yes, you can delete it

By the way, this version of the EA takes quite sometime during training and so I guess it is not required for retraining everyday..but still I will decide after I see the drawdown and consecutive losses..

Also, do you have any idea as how to feed random values into RDF for training? I mean I don't want to feed anything from past data to RDF. I don't know whether it is possible in this version of the RDF or not and so before I proceed I just thought to check with you

Here is my algo which I am trying to implement:

1.No optimization

2.No past data training

3.No indicators or No price feed and nothing

4.Just using random candle simulations

5.I mean in step 4, one agent will create random candle lengths based on a maximum value set for the candle length and candle directions(BUY or SELL) and the second agent will

take those values of candle length and the direction as input to RDF at each step and place BUY and SELL orders and will try to maximise the reward. In this case, the reward will be the summation or subtraction of candle lengths from starting point.

6.So we can just train this model for millions of such candles with different variations in which RDF will learn to maximise the profits directly and we can just keep the training ON..

If it is feasible, the only hurdle is just the computing power as how many simulations and checks we can do..

By the way, I have the sample code for random candle generation and candle direction(BUY or SELL), but don't have visual mode, but only values. So I am looking for as how to feed those values to RDF for training and then, it should use that model while trading 

 

Я смотрю и фокусник уже к буржуям метит. Тогда вот Вам и мой английски.

Vi DURAKI

ну как??? Звучит? :-)))))))

 
mytarmailS:

уровни....

уровни работают....

еще немного допилю, а потом сверху можно и нейронку вешать как фильтр )

нейронку?

Мне кажется что местный форум (этой ветки касаемо) надеется на машинное обучение как на то, что спасет их от необходимости думать

Если у Вас уже есть визуально хороший результат, неужели Вы не сможете составить алгоритм по принципу "если то" ?

 
Mihail Marchukajtes:

Я смотрю и фокусник уже к буржуям метит. Тогда вот Вам и мой английски.

Vi DURAKI

ну как??? Звучит? :-)))))))

неее, не прокатит, нужно на международном: Ю ЛУЗЕР!!!  

))))

 
FxTrader562:

By the way, this version of the EA takes quite sometime during training and so I guess it is not required for retraining everyday..but still I will decide after I see the drawdown and consecutive losses..

Also, do you have any idea as how to feed random values into RDF for training? I mean I don't want to feed anything from past data to RDF. I don't know whether it is possible in this version of the RDF or not and so before I proceed I just thought to check with you

Here is my algo which I am trying to implement:

1.No optimization

2.No past data training

3.No indicators or No price feed and nothing

4.Just using random candle simulations

5.I mean in step 4, one agent will create random candle lengths based on a maximum value set for the candle length and candle directions(BUY or SELL) and the second agent will

take those values of candle length and the direction as input to RDF at each step and place BUY and SELL orders and will try to maximise the reward. In this case, the reward will be the summation or subtraction of candle lengths from starting point.

6.So we can just train this model for millions of such candles with different variations in which RDF will learn to maximise the profits directly and we can just keep the training ON..

If it is feasible, the only hurdle is just the computing power as how many simulations and checks we can do..

By the way, I have the sample code for random candle generation and candle direction(BUY or SELL), but don't have visual mode, but only values. So I am looking for as how to feed those values to RDF for training and then, it should use that model while trading 

You can just feed random price values instead close prices in learning stage

void calcSignal()
  {
   sig1=0;
   
   if(learn==false)   
   for(int i=0;i<ArraySize(ag1.agent);i++)
     {   
      CopyClose(_Symbol,0,0,ArraySize(ag1.agent[i].inpVector),ag1.agent[i].inpVector);
      normalizeArrays(ag1.agent[i].inpVector);
     }
     
   else
   for(int i=0;i<ArraySize(ag1.agent);i++)
     {
      for(int l=0;i<ArraySize(ag1.agent[l].inpVector);l++)
        {
         double pr[];
         CopyClose(_Symbol,0,rand(),1,pr);
         ag1.agent[i].inpVector[l] = pr[0];
        }   
   
      normalizeArrays(ag1.agent[i].inpVector);
     }
   sig1=ag1.getTradeSignal();
  }
something like this
 
Maxim Dmitrievsky :

You can just buy a random price

Ok, but I am confused about some of the codes that you used before.

CRLAgents * ag1 = new CRLAgents ("RlExp1iter", 5,100,50, regularize, learn);

and this one:

CopyClose (_Symbol, 0,0,100, ag1.agent [i] .inpVector);

So if we have to use 100 candles in copyclose. right? I mean both candles and features are same. right?

Or we can use different values ​​for features and candles?

By the way, I am trying to use 1000 features and the training is on for 1 hour now ..
 
FxTrader562:

Ok, but I am confused some of the codes before I implement something of my own.The below code:

CRLAgents *ag1=new CRLAgents("RlExp1iter",5,100,50,regularize,learn);

and this one:

CopyClose(_Symbol,0,0,100,ag1.agent[i].inpVector);

So if we have to use 100 features in the above code means we have to use 100 candles in copyclose. right?

Or we can use different values for features and candles?

sure you can use different values for each predictor, its just a simple example, every close value = 1 difeerent predictor value

 
FxTrader562:
By the way, I am trying to use 1000 features and the training is on for 1 hour now..

you can set only 1 agent CRLAgents *ag1=new CRLAgents("RlExp1iter",1,100,50,regularize,learn);


and in library set #define  _models 1

so it will be fast

 
Maxim Dmitrievsky:

sure you can use different values for each predictor, its just a simple example, every close value = 1 difeerent predictor value

So the number 100 or 1000 or 500 etc has to be same in both the codes copyclose and in declaration.right?

1...106010611062106310641065106610671068106910701071107210731074...3494
Новый комментарий
Причина обращения: