forked from dng/NN_in_Trading
1198 lines
44 KiB
MQL5
1198 lines
44 KiB
MQL5
//+------------------------------------------------------------------+
|
|
//| Trajectory.mqh |
|
|
//| Copyright 2026, DNG |
|
|
//| https://www.mql5.com |
|
|
//+------------------------------------------------------------------+
|
|
//| |
|
|
//| Trajectory state, network topology and RAG memory helpers |
|
|
//| for the VLADriver-RAG Expert Advisor. |
|
|
//+------------------------------------------------------------------+
|
|
#include "..\NeuroNet_DNG\NeuroNet.mqh"
|
|
#include <Trade\Trade.mqh>
|
|
#include <Trade\SymbolInfo.mqh>
|
|
#include <Indicators\Oscilators.mqh>
|
|
input group "---- Indicators ----"
|
|
input ENUM_TIMEFRAMES TimeFrame = PERIOD_H1;
|
|
input group "---- RSI ----"
|
|
input int RSIPeriod = 14; //Period
|
|
input ENUM_APPLIED_PRICE RSIPrice = PRICE_CLOSE; //Applied price
|
|
input group "---- CCI ----"
|
|
input int CCIPeriod = 14; //Period
|
|
input ENUM_APPLIED_PRICE CCIPrice = PRICE_TYPICAL; //Applied price
|
|
input group "---- ATR ----"
|
|
input int ATRPeriod = 14; //Period
|
|
input group "---- MACD ----"
|
|
input int FastPeriod = 12; //Fast
|
|
input int SlowPeriod = 26; //Slow
|
|
input int SignalPeriod = 9; //Signal
|
|
input ENUM_APPLIED_PRICE MACDPrice = PRICE_CLOSE; //Applied price
|
|
input group "---- RAG runtime ----"
|
|
// Completed records staged before the library publishes a new immutable
|
|
// snapshot. It is collection policy, never part of layer topology.
|
|
input uint OnlineMemorySize = 256;
|
|
// Domain facts supplied with completed trading records for normalization.
|
|
input double MemoryRiskBudgetFraction = 0.01;
|
|
input double MemoryVolatilityRange = 10.0;
|
|
int iLatentLayer = -1;
|
|
#define HistoryBars 5
|
|
#define BarDescr 9 //Elements for 1 bar description
|
|
#define AccountDescr 13 //Account description
|
|
#define NActions 6 //Number of possible Actions
|
|
#define NRewards 1 //Number of rewards
|
|
#define NForecast 12 //Number of forecast
|
|
#define EtalonBalance 1e4
|
|
#define BatchSize 1e+5
|
|
#define EmbeddingSize 16
|
|
#define DiscFactor 0.5f
|
|
#define FileName "VLADriverRAG"
|
|
#define ActorGraphVersion 3
|
|
#define CriticCheckpointFile "Crt.nnw"
|
|
#define LatentCount 64
|
|
#define LatentLayer iLatentLayer
|
|
#define StateScenarioLayer 7
|
|
// The Critic attends to the full causal RankTCM output [BarDescr, EmbeddingSize].
|
|
// The Actor/RAG path uses the separately pooled causal scenario embedding.
|
|
#define StateTokenLayer 4
|
|
#define ForecastTokenDim (EmbeddingSize + 1)
|
|
#define MaxSL 1000
|
|
#define MaxTP 1000
|
|
#define ActorUpdate 5
|
|
#define TargetUpdate 24*5
|
|
#define tau 0.9f
|
|
#define NHeads 4
|
|
#define NExperts 5
|
|
#define RAGScenarioCentroids 512 //CLayerDescription capacity
|
|
#define RAGActionCentroids 24 //per-scenario CLayerDescription capacity
|
|
#define NScenarios 3 //Account, position and time primary candidates
|
|
#define ActorMPITopK 3 //Must not exceed the three Actor primary candidates
|
|
#define TopK 5 //Critic candidate selection remains independent from Actor MPI
|
|
#define Quantiles 8
|
|
// CogDriverData requires its state stack to hold at least ten quantile slots.
|
|
// This is a MarketEncoder constraint, independent of Actor primary candidates.
|
|
#define StackSize (10*Quantiles)
|
|
#define Blocks 24
|
|
CSymbolInfo Symb;
|
|
CTrade Trade;
|
|
MqlRates Rates[];
|
|
CiRSI RSI;
|
|
CiCCI CCI;
|
|
CiATR ATR;
|
|
CiMACD MACD;
|
|
struct SState
|
|
{
|
|
float state[HistoryBars * BarDescr];
|
|
float account[AccountDescr - 4];
|
|
float action[NActions];
|
|
float rewards[NRewards];
|
|
SState(void);
|
|
bool Save(int file_handle);
|
|
bool Load(int file_handle);
|
|
// Overloading.
|
|
void operator=(const SState &obj)
|
|
{
|
|
ArrayCopy(state, obj.state);
|
|
ArrayCopy(account, obj.account);
|
|
ArrayCopy(action, obj.action);
|
|
ArrayCopy(rewards, obj.rewards);
|
|
}
|
|
};
|
|
//| SState implementation. |
|
|
//+------------------------------------------------------------------+
|
|
SState::SState(void)
|
|
{
|
|
ArrayInitialize(state, 0);
|
|
ArrayInitialize(account, 0);
|
|
ArrayInitialize(action, 0);
|
|
ArrayInitialize(rewards, 0);
|
|
}
|
|
//| Save implementation. |
|
|
//+------------------------------------------------------------------+
|
|
bool SState::Save(int file_handle)
|
|
{
|
|
if(file_handle == INVALID_HANDLE)
|
|
ReturnFalse;
|
|
int total = ArraySize(state);
|
|
if(FileWriteInteger(file_handle, total) < sizeof(int))
|
|
ReturnFalse;
|
|
for(int i = 0; i < total; i++)
|
|
if(FileWriteFloat(file_handle, state[i]) < sizeof(float))
|
|
ReturnFalse;
|
|
total = ArraySize(account);
|
|
if(FileWriteInteger(file_handle, total) < sizeof(int))
|
|
ReturnFalse;
|
|
for(int i = 0; i < total; i++)
|
|
if(FileWriteFloat(file_handle, account[i]) < sizeof(float))
|
|
ReturnFalse;
|
|
total = ArraySize(action);
|
|
if(FileWriteInteger(file_handle, total) < sizeof(int))
|
|
ReturnFalse;
|
|
for(int i = 0; i < total; i++)
|
|
if(FileWriteFloat(file_handle, action[i]) < sizeof(float))
|
|
ReturnFalse;
|
|
total = ArraySize(rewards);
|
|
if(FileWriteInteger(file_handle, total) < sizeof(int))
|
|
ReturnFalse;
|
|
for(int i = 0; i < total; i++)
|
|
if(FileWriteFloat(file_handle, rewards[i]) < sizeof(float))
|
|
ReturnFalse;
|
|
return true;
|
|
}
|
|
//| Load implementation. |
|
|
//+------------------------------------------------------------------+
|
|
bool SState::Load(int file_handle)
|
|
{
|
|
if(file_handle == INVALID_HANDLE)
|
|
ReturnFalse;
|
|
if(FileIsEnding(file_handle))
|
|
ReturnFalse;
|
|
int total = FileReadInteger(file_handle);
|
|
if(total != ArraySize(state))
|
|
ReturnFalse;
|
|
for(int i = 0; i < total; i++)
|
|
{
|
|
if(FileIsEnding(file_handle))
|
|
ReturnFalse;
|
|
state[i] = FileReadFloat(file_handle);
|
|
}
|
|
total = FileReadInteger(file_handle);
|
|
if(total != ArraySize(account))
|
|
ReturnFalse;
|
|
for(int i = 0; i < total; i++)
|
|
{
|
|
if(FileIsEnding(file_handle))
|
|
ReturnFalse;
|
|
account[i] = FileReadFloat(file_handle);
|
|
}
|
|
total = FileReadInteger(file_handle);
|
|
if(total != ArraySize(action))
|
|
ReturnFalse;
|
|
for(int i = 0; i < total; i++)
|
|
{
|
|
if(FileIsEnding(file_handle))
|
|
ReturnFalse;
|
|
action[i] = MathMin(MathMax(FileReadFloat(file_handle), 0), 1);
|
|
}
|
|
total = FileReadInteger(file_handle);
|
|
if(total != ArraySize(rewards))
|
|
ReturnFalse;
|
|
for(int i = 0; i < total; i++)
|
|
{
|
|
if(FileIsEnding(file_handle))
|
|
ReturnFalse;
|
|
rewards[i] = FileReadFloat(file_handle);
|
|
}
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| CreateStateDescriptions. |
|
|
//+------------------------------------------------------------------+
|
|
bool CreateStateDescriptions(CArrayObj *&encoder,
|
|
CArrayObj *&decoder
|
|
)
|
|
{
|
|
CLayerDescription *descr;
|
|
if(!encoder)
|
|
{
|
|
encoder = new CArrayObj();
|
|
if(!encoder)
|
|
ReturnFalse;
|
|
}
|
|
if(!decoder)
|
|
{
|
|
decoder = new CArrayObj();
|
|
if(!decoder)
|
|
ReturnFalse;
|
|
}
|
|
// State Encoder.
|
|
encoder.Clear();
|
|
// Input layer.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBaseOCL;
|
|
uint prev_count = descr.count = HistoryBars * BarDescr;
|
|
descr.activation = None;
|
|
descr.optimization = ADAM;
|
|
if(!encoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 1.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBatchNormOCL;
|
|
descr.count = prev_count;
|
|
descr.batch = 1e4;
|
|
descr.activation = None;
|
|
descr.optimization = ADAM;
|
|
if(!encoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 2.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronDropoutOCL;
|
|
descr.count = prev_count;
|
|
descr.probability = 0.1f;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!encoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 3.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronCogDriverData;
|
|
descr.window = HistoryBars;
|
|
descr.count = BarDescr;
|
|
{
|
|
uint temp[] = {StackSize, StackSize, Quantiles};
|
|
if(ArrayCopy(descr.units, temp, 0, 0, temp.Size()) < int(temp.Size()))
|
|
ReturnFalse;
|
|
}
|
|
descr.probability = 1.0f;
|
|
descr.activation = None;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!encoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 4.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronCogDriverRankTCM;
|
|
descr.window = HistoryBars * (2 * Quantiles + 1);
|
|
descr.count = EmbeddingSize;
|
|
descr.variables = BarDescr;
|
|
{
|
|
uint temp[] = {StackSize, NHeads};
|
|
if(ArrayCopy(descr.units, temp, 0, 0, temp.Size()) < int(temp.Size()))
|
|
ReturnFalse;
|
|
}
|
|
descr.activation = None;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!encoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 5.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronSpikeConv;
|
|
descr.window = EmbeddingSize;
|
|
descr.step = EmbeddingSize;
|
|
descr.window_out = EmbeddingSize / 2;
|
|
descr.count = 1;
|
|
descr.variables = BarDescr;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!encoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 6.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBaseOCL;
|
|
descr.count = EmbeddingSize;
|
|
descr.batch = BatchSize;
|
|
descr.activation = None;
|
|
descr.optimization = ADAM;
|
|
if(!encoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 7.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronSpikeActivation;
|
|
descr.count = EmbeddingSize;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!encoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// Forecast Decoder.
|
|
decoder.Clear();
|
|
// Input layer.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBaseOCL;
|
|
prev_count = descr.count = EmbeddingSize;
|
|
descr.activation = None;
|
|
descr.optimization = ADAM;
|
|
if(!decoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 1.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronConvOCL;
|
|
descr.count = 1;
|
|
descr.window = EmbeddingSize;
|
|
descr.step = EmbeddingSize;
|
|
descr.window_out = NForecast * EmbeddingSize;
|
|
descr.activation = None;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!decoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 2.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronConvOCL;
|
|
descr.count = NForecast;
|
|
descr.window = EmbeddingSize;
|
|
descr.step = EmbeddingSize;
|
|
descr.window_out = 2 * EmbeddingSize;
|
|
descr.activation = GELU;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!decoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 3.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronConvOCL;
|
|
descr.count = NForecast;
|
|
descr.window = 2 * EmbeddingSize;
|
|
descr.step = 2 * EmbeddingSize;
|
|
descr.window_out = EmbeddingSize;
|
|
descr.activation = GELU;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!decoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 4.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronConvOCL;
|
|
descr.count = NForecast;
|
|
descr.window = EmbeddingSize;
|
|
descr.step = EmbeddingSize;
|
|
descr.window_out = BarDescr;
|
|
descr.activation = None;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!decoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 5.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBatchNormOCL;
|
|
descr.count = NForecast * BarDescr;
|
|
descr.batch = 1e4;
|
|
descr.activation = None;
|
|
descr.optimization = ADAM;
|
|
if(!decoder.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| CreateDescriptions. |
|
|
//+------------------------------------------------------------------+
|
|
bool CreateDescriptions(CArrayObj *&actor,
|
|
CArrayObj *&critic
|
|
)
|
|
{
|
|
CLayerDescription *descr;
|
|
if(!actor)
|
|
{
|
|
actor = new CArrayObj();
|
|
if(!actor)
|
|
ReturnFalse;
|
|
}
|
|
if(!critic)
|
|
{
|
|
critic = new CArrayObj();
|
|
if(!critic)
|
|
ReturnFalse;
|
|
}
|
|
// Actor.
|
|
actor.Clear();
|
|
// Input layer.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBaseOCL;
|
|
uint prev_count = descr.count = AccountDescr;
|
|
descr.activation = None;
|
|
descr.optimization = ADAM;
|
|
if(!actor.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
iLatentLayer = 0;
|
|
// layer 1: account/position/time candidate projection.
|
|
// This learned FC projection turns the compact AccountDescr into the three
|
|
// primary MPI candidates; market and RAG remain auxiliary contexts.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBaseOCL;
|
|
descr.count = NScenarios * EmbeddingSize;
|
|
descr.activation = GELU;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!actor.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 2: MPI primary candidates, market second input and bound RAG third input.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronVLADriverRAGMPI;
|
|
descr.window = EmbeddingSize;
|
|
descr.count = StackSize;
|
|
{
|
|
// CNeuronVLADriverRAGMPI consumes a single frozen ScenarioEmbedding
|
|
// through its second input and learns NHeads market K/V tokens locally.
|
|
uint temp[] = {NScenarios, EmbeddingSize, NHeads};
|
|
if(ArrayCopy(descr.units, temp, 0, 0, temp.Size()) < int(temp.Size()))
|
|
ReturnFalse;
|
|
}
|
|
descr.probability = ActorMPITopK;
|
|
descr.step = NHeads;
|
|
descr.window_out = EmbeddingSize / NHeads;
|
|
// One local action-centroid set is gathered for each selected scenario.
|
|
// The RAG-memory descriptor below uses the same Top-K, so binding is exact.
|
|
descr.layers = ActorMPITopK * RAGActionCentroids;
|
|
descr.variables = (NActions + 2);
|
|
descr.activation = None;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!actor.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
iLatentLayer = 2;
|
|
// layer 3.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronSpikeConvBlock;
|
|
descr.count = 1;
|
|
descr.window = EmbeddingSize;
|
|
descr.step = EmbeddingSize;
|
|
descr.window_out = EmbeddingSize;
|
|
descr.variables = 1;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!actor.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 4.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBaseOCL;
|
|
descr.count = NActions;
|
|
descr.optimization = ADAM;
|
|
if(!actor.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 5.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronConvOCL;
|
|
descr.count = NActions / 3;
|
|
descr.window = 3;
|
|
descr.step = 3;
|
|
descr.window_out = 3;
|
|
descr.activation = SIGMOID;
|
|
descr.optimization = ADAM;
|
|
if(!actor.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// Critic.
|
|
critic.Clear();
|
|
// Input layer.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBaseOCL;
|
|
descr.count = NActions;
|
|
descr.activation = None;
|
|
descr.optimization = ADAM;
|
|
if(!critic.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 1.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronMHCrossFAT;
|
|
{
|
|
uint temp[] = {3, // Inputs window
|
|
EmbeddingSize, // Key Dimension
|
|
EmbeddingSize, // Cross window
|
|
EmbeddingSize // Embedding size
|
|
};
|
|
if(ArrayCopy(descr.windows, temp) < (int)temp.Size())
|
|
ReturnFalse;
|
|
}
|
|
{
|
|
uint temp[] = {NActions / 3, // Query units
|
|
BarDescr // Cross units
|
|
};
|
|
if(ArrayCopy(descr.units, temp) < (int)temp.Size())
|
|
ReturnFalse;
|
|
}
|
|
descr.step = NHeads; // Heads
|
|
descr.batch = 1e4;
|
|
descr.layers = NExperts; // Candidates
|
|
descr.variables = TopK; // Top-K
|
|
descr.activation = None;
|
|
descr.optimization = ADAM;
|
|
if(!critic.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 2.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronSpikeConvBlock;
|
|
descr.count = NActions / 3;
|
|
descr.window = 3;
|
|
descr.step = 3;
|
|
descr.window_out = EmbeddingSize;
|
|
descr.variables = 1;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!critic.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 4.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronSpikeConvBlock;
|
|
descr.count = NActions / 3;
|
|
descr.window = EmbeddingSize;
|
|
descr.step = EmbeddingSize;
|
|
descr.window_out = EmbeddingSize;
|
|
descr.variables = 1;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!critic.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 5.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBaseOCL;
|
|
prev_count = descr.count = LatentCount;
|
|
descr.activation = SIGMOID;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!critic.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
// layer 6.
|
|
if(!(descr = new CLayerDescription()))
|
|
DeleteObjAndFalse(descr);
|
|
descr.type = defNeuronBaseOCL;
|
|
prev_count = descr.count = NRewards;
|
|
descr.activation = None;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
if(!critic.Add(descr))
|
|
DeleteObjAndFalse(descr);
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Shared RAG-memory layer shape. The memory remains library-owned; |
|
|
//+------------------------------------------------------------------+
|
|
//| Implements CreateRAGMemoryDescription |
|
|
bool CreateRAGMemoryDescription(CLayerDescription &descr)
|
|
{
|
|
descr.type = defNeuronRAGMemory;
|
|
descr.window = EmbeddingSize;
|
|
descr.window_out = NActions;
|
|
descr.count = NActions + 2;
|
|
descr.layers = ActorMPITopK;
|
|
descr.activation = None;
|
|
descr.batch = BatchSize;
|
|
descr.optimization = ADAM;
|
|
uint scenario_capacity[] = {RAGScenarioCentroids};
|
|
uint action_capacity[] = {RAGActionCentroids};
|
|
if(ArrayCopy(descr.units, scenario_capacity, 0, 0, scenario_capacity.Size()) < int(scenario_capacity.Size()) ||
|
|
ArrayCopy(descr.heads, action_capacity, 0, 0, action_capacity.Size()) < int(action_capacity.Size()))
|
|
ReturnFalse;
|
|
return true;
|
|
}
|
|
#ifndef Study
|
|
//| IsNewBar implementation. |
|
|
//+------------------------------------------------------------------+
|
|
bool IsNewBar(void)
|
|
{
|
|
static datetime last_bar = 0;
|
|
if(last_bar >= iTime(Symb.Name(), TimeFrame, 0))
|
|
return false;
|
|
last_bar = iTime(Symb.Name(), TimeFrame, 0);
|
|
return true;
|
|
}
|
|
//| CloseByDirection implementation. |
|
|
//+------------------------------------------------------------------+
|
|
bool CloseByDirection(ENUM_POSITION_TYPE type)
|
|
{
|
|
int total = PositionsTotal();
|
|
bool result = true;
|
|
for(int i = total - 1; i >= 0; i--)
|
|
{
|
|
if(PositionGetSymbol(i) != Symb.Name())
|
|
continue;
|
|
if(PositionGetInteger(POSITION_TYPE) != type)
|
|
continue;
|
|
result = (Trade.PositionClose(PositionGetInteger(POSITION_TICKET)) && result);
|
|
}
|
|
return result;
|
|
}
|
|
//| TrailPosition implementation. |
|
|
//+------------------------------------------------------------------+
|
|
bool TrailPosition(ENUM_POSITION_TYPE type, double sl, double tp)
|
|
{
|
|
int total = PositionsTotal();
|
|
bool result = true;
|
|
datetime time = TimeCurrent() - 5 * PeriodSeconds(TimeFrame);
|
|
for(int i = 0; i < total; i++)
|
|
{
|
|
if(PositionGetSymbol(i) != Symb.Name())
|
|
continue;
|
|
if(PositionGetInteger(POSITION_TYPE) != type)
|
|
continue;
|
|
if(PositionGetInteger(POSITION_TIME_UPDATE) > time)
|
|
continue;
|
|
bool modify = false;
|
|
double psl = PositionGetDouble(POSITION_SL);
|
|
double ptp = PositionGetDouble(POSITION_TP);
|
|
switch(type)
|
|
{
|
|
case POSITION_TYPE_BUY:
|
|
if((sl - psl) >= Symb.Point())
|
|
{
|
|
psl = sl;
|
|
modify = true;
|
|
}
|
|
if(MathAbs(tp - ptp) >= Symb.Point())
|
|
{
|
|
ptp = tp;
|
|
modify = true;
|
|
}
|
|
break;
|
|
case POSITION_TYPE_SELL:
|
|
if((psl - sl) >= Symb.Point())
|
|
{
|
|
psl = sl;
|
|
modify = true;
|
|
}
|
|
if(MathAbs(tp - ptp) >= Symb.Point())
|
|
{
|
|
ptp = tp;
|
|
modify = true;
|
|
}
|
|
break;
|
|
}
|
|
if(modify)
|
|
result = (Trade.PositionModify(PositionGetInteger(POSITION_TICKET), psl, ptp) && result);
|
|
}
|
|
return result;
|
|
}
|
|
//| ClosePartial implementation. |
|
|
//+------------------------------------------------------------------+
|
|
bool ClosePartial(ENUM_POSITION_TYPE type, double value)
|
|
{
|
|
if(value <= 0)
|
|
return true;
|
|
for(int i = 0; (i < PositionsTotal() && value > 0); i++)
|
|
{
|
|
if(PositionGetSymbol(i) != Symb.Name())
|
|
continue;
|
|
if(PositionGetInteger(POSITION_TYPE) != type)
|
|
continue;
|
|
double pvalue = PositionGetDouble(POSITION_VOLUME);
|
|
if(pvalue <= value)
|
|
{
|
|
if(Trade.PositionClose(PositionGetInteger(POSITION_TICKET)))
|
|
{
|
|
value -= pvalue;
|
|
i--;
|
|
}
|
|
}
|
|
else
|
|
{
|
|
if(Trade.PositionClosePartial(PositionGetInteger(POSITION_TICKET), value))
|
|
value = 0;
|
|
}
|
|
}
|
|
return (value <= 0);
|
|
}
|
|
#endif
|
|
//| Implements CreateBuffers |
|
|
bool CreateBuffers(const int start_bar, CBufferFloat* state, CBufferFloat *time, CBufferFloat* forecast)
|
|
{
|
|
int total_bars = (start_bar + HistoryBars + (!!forecast ? NForecast : 0));
|
|
if(!state || !time || start_bar < 0 ||
|
|
total_bars > int(Rates.Size()))
|
|
ReturnFalse;
|
|
matrix<float> mState = matrix<float>::Zeros(BarDescr, HistoryBars);
|
|
vector<float> vForecast = vector<float>::Zeros(NForecast * BarDescr);
|
|
time.Clear();
|
|
time.Reserve(HistoryBars);
|
|
int bar = start_bar + (!!forecast ? NForecast : 0);
|
|
for(int b = 0; b < (int)HistoryBars; b++)
|
|
{
|
|
float open = (float)Rates[b + bar].open;
|
|
float rsi = (float)RSI.Main(b + bar);
|
|
float cci = (float)CCI.Main(b + bar);
|
|
float atr = (float)ATR.Main(b + bar);
|
|
float macd = (float)MACD.Main(b + bar);
|
|
float sign = (float)MACD.Signal(b + bar);
|
|
if(rsi == EMPTY_VALUE || cci == EMPTY_VALUE || atr == EMPTY_VALUE || macd == EMPTY_VALUE || sign == EMPTY_VALUE)
|
|
ReturnFalse;
|
|
mState[0, b] = (float)(Rates[b + bar].close - open);
|
|
mState[1, b] = (float)(Rates[b + bar].high - open);
|
|
mState[2, b] = (float)(Rates[b + bar].low - open);
|
|
mState[3, b] = (float)(Rates[b + bar].tick_volume / 1000.0f);
|
|
mState[4, b] = rsi;
|
|
mState[5, b] = cci;
|
|
mState[6, b] = atr;
|
|
mState[7, b] = macd;
|
|
mState[8, b] = sign;
|
|
if(!time.Add(float(Rates[b + bar].time)))
|
|
ReturnFalse;
|
|
}
|
|
if(!state.AssignArray(mState))
|
|
ReturnFalse;
|
|
if(time.GetIndex() >= 0)
|
|
if(!time.BufferWrite())
|
|
ReturnFalse;
|
|
if(!forecast)
|
|
return true;
|
|
for(int b = 1; b <= (int)NForecast; b++)
|
|
{
|
|
float open = (float)Rates[bar - b].open;
|
|
float rsi = (float)RSI.Main(bar - b);
|
|
float cci = (float)CCI.Main(bar - b);
|
|
float atr = (float)ATR.Main(bar - b);
|
|
float macd = (float)MACD.Main(bar - b);
|
|
float sign = (float)MACD.Signal(bar - b);
|
|
if(rsi == EMPTY_VALUE || cci == EMPTY_VALUE || atr == EMPTY_VALUE || macd == EMPTY_VALUE || sign == EMPTY_VALUE)
|
|
ReturnFalse;
|
|
int shift = (NForecast - b) * BarDescr;
|
|
vForecast[shift] = (float)(Rates[bar - b].close - open);
|
|
vForecast[shift + 1] = (float)(Rates[bar - b].high - open);
|
|
vForecast[shift + 2] = (float)(Rates[bar - b].low - open);
|
|
vForecast[shift + 3] = (float)(Rates[bar - b].tick_volume / 1000.0f);
|
|
vForecast[shift + 4] = rsi;
|
|
vForecast[shift + 5] = cci;
|
|
vForecast[shift + 6] = atr;
|
|
vForecast[shift + 7] = macd;
|
|
vForecast[shift + 8] = sign;
|
|
}
|
|
if(!forecast.AssignArray(vForecast))
|
|
ReturnFalse;
|
|
return true;
|
|
}
|
|
enum ENUM_MEMORY_ACTION_NORMALIZATION_STAGE
|
|
{
|
|
MEMORY_ACTION_NORMALIZATION_OK = 0,
|
|
MEMORY_ACTION_NORMALIZATION_INPUT,
|
|
MEMORY_ACTION_NORMALIZATION_MARKET,
|
|
MEMORY_ACTION_NORMALIZATION_VOLATILITY,
|
|
MEMORY_ACTION_NORMALIZATION_DESCRIPTOR,
|
|
MEMORY_ACTION_NORMALIZATION_ZERO
|
|
};
|
|
//| MemoryActionNormalizationStageName implementation. |
|
|
//| Returns a human-readable name for each normalization stage. |
|
|
//+------------------------------------------------------------------+
|
|
string MemoryActionNormalizationStageName(const ENUM_MEMORY_ACTION_NORMALIZATION_STAGE stage)
|
|
{
|
|
switch(stage)
|
|
{
|
|
case MEMORY_ACTION_NORMALIZATION_INPUT:
|
|
return "input";
|
|
case MEMORY_ACTION_NORMALIZATION_MARKET:
|
|
return "market";
|
|
case MEMORY_ACTION_NORMALIZATION_VOLATILITY:
|
|
return "volatility";
|
|
case MEMORY_ACTION_NORMALIZATION_DESCRIPTOR:
|
|
return "descriptor";
|
|
case MEMORY_ACTION_NORMALIZATION_ZERO:
|
|
return "zero";
|
|
default:
|
|
return "ok";
|
|
}
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| NormalizeMemoryAction (with diagnostics). |
|
|
//+------------------------------------------------------------------+
|
|
bool NormalizeMemoryAction(CNeuronRAGMemory &memory, CBufferFloat *action, const uint start_position,
|
|
const double equity, float &destination[],
|
|
ENUM_MEMORY_ACTION_NORMALIZATION_STAGE &stage)
|
|
{
|
|
stage = MEMORY_ACTION_NORMALIZATION_INPUT;
|
|
ArrayInitialize(destination, 0.0f);
|
|
if(action == NULL || action.Total() < NActions || ArraySize(destination) < NActions ||
|
|
equity <= 0 || start_position >= Rates.Size())
|
|
ReturnFalse;
|
|
stage = MEMORY_ACTION_NORMALIZATION_MARKET;
|
|
if(!Symb.RefreshRates())
|
|
ReturnFalse;
|
|
const double volatility = ATR.Main(start_position);
|
|
const double point_cost = Symb.TickValue() / Symb.TickSize();
|
|
if(point_cost <= 0 || Symb.Point() <= 0)
|
|
ReturnFalse;
|
|
stage = MEMORY_ACTION_NORMALIZATION_VOLATILITY;
|
|
if(volatility == EMPTY_VALUE || volatility <= 0)
|
|
ReturnFalse;
|
|
float raw[NActions];
|
|
for(int i = 0; i < NActions; i++)
|
|
raw[i] = action[i];
|
|
const double risk_budget = equity * MemoryRiskBudgetFraction;
|
|
stage = MEMORY_ACTION_NORMALIZATION_DESCRIPTOR;
|
|
if(!memory.NormalizeActionDescriptor(raw, equity, risk_budget, point_cost, Symb.Point(),
|
|
volatility, MemoryVolatilityRange, MaxTP, MaxSL, destination))
|
|
ReturnFalse;
|
|
double action_square = 0.0;
|
|
for(int i = 0; i < NActions; i++)
|
|
action_square += destination[i] * destination[i];
|
|
stage = MEMORY_ACTION_NORMALIZATION_ZERO;
|
|
if(action_square <= DBL_EPSILON)
|
|
ReturnFalse;
|
|
stage = MEMORY_ACTION_NORMALIZATION_OK;
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| Preserve the existing callers that do not need diagnostics. |
|
|
//+------------------------------------------------------------------+
|
|
bool NormalizeMemoryAction(CNeuronRAGMemory &memory, CBufferFloat *action, const uint start_position,
|
|
const double equity, float &destination[])
|
|
{
|
|
ENUM_MEMORY_ACTION_NORMALIZATION_STAGE stage;
|
|
return NormalizeMemoryAction(memory, action, start_position, equity, destination, stage);
|
|
}
|
|
const vector<float> SampleAccount(CBufferFloat *state, datetime time, double max_balance, double min_balance = 0)
|
|
{
|
|
vector<float> result = vector<float>::Zeros(AccountDescr);
|
|
if(!state)
|
|
return result;
|
|
double marg = 0;
|
|
if(!Symb.RefreshRates() || !OrderCalcMargin(ORDER_TYPE_BUY, Symb.Name(), 1, Symb.Ask(), marg))
|
|
return result;
|
|
double buy_lot = 0, sell_lot = 0, profit = 0;
|
|
double deal = 0;//MathRand() / (32767 * 0.5) - 1;
|
|
double multiplyer = 1.0 / (60.0 * 60.0 * 10.0);
|
|
double balance = (max_balance - min_balance) * MathRand() / 32767.0 + min_balance;
|
|
double prev_balance = balance;
|
|
double equity = balance;
|
|
double prev_equity = balance;
|
|
double position_discount = 0;
|
|
if(deal > 0)
|
|
{
|
|
double lot = balance / (2.0 * marg) * deal;
|
|
if(lot >= Symb.LotsMin())
|
|
buy_lot = MathMin(int((lot - Symb.LotsMin()) / Symb.LotsStep()) * Symb.LotsStep() + Symb.LotsMin(), 1.0);
|
|
}
|
|
else
|
|
if(deal < 0)
|
|
{
|
|
double lot = MathAbs(balance / (2.0 * marg) * deal);
|
|
if(lot >= Symb.LotsMin())
|
|
sell_lot = MathMin(int((lot - Symb.LotsMin()) / Symb.LotsStep()) * Symb.LotsStep() + Symb.LotsMin(), 1.0);
|
|
}
|
|
else
|
|
prev_balance += (MathRand() / (2.0 * 32767.0) - 0.25) * balance;
|
|
if(sell_lot > 0 || buy_lot > 0)
|
|
{
|
|
int pos_open = int(MathRand() / 32767.0 * (state.Total() / BarDescr - 1));
|
|
for(int i = 0; i <= pos_open; i++)
|
|
{
|
|
profit += state.At(i * BarDescr) / Symb.TickSize() * Symb.TickValue();
|
|
if(((buy_lot > 0 && profit < 0) ||
|
|
(sell_lot > 0 && profit > 0))
|
|
&& MathAbs(profit * (buy_lot - sell_lot)) > balance / 2)
|
|
{
|
|
pos_open = i;
|
|
break;
|
|
}
|
|
}
|
|
profit *= buy_lot - sell_lot;
|
|
equity += profit;
|
|
prev_equity = equity - state.At(0) / Symb.TickSize() * Symb.TickValue() * (buy_lot - sell_lot);
|
|
position_discount = pos_open * PeriodSeconds(TimeFrame) * multiplyer * MathAbs(profit);
|
|
}
|
|
result[0] = float(balance / EtalonBalance);
|
|
result[1] = float((balance - prev_balance) / prev_balance);
|
|
result[2] = float(equity / prev_balance);
|
|
result[3] = float((equity - prev_equity) / prev_equity);
|
|
result[4] = float(buy_lot);
|
|
result[5] = float(sell_lot);
|
|
result[6] = float(buy_lot * profit / prev_balance);
|
|
result[7] = float(sell_lot * profit / prev_balance);
|
|
result[8] = float(position_discount / prev_balance);
|
|
double x = time / (double)(D'2024.01.01' - D'2023.01.01');
|
|
result[9] = float(MathSin(x != 0 ? 2.0 * M_PI * x : 0));
|
|
x = time / (double)PeriodSeconds(PERIOD_MN1);
|
|
result[10] = float(MathCos(x != 0 ? 2.0 * M_PI * x : 0));
|
|
x = time / (double)PeriodSeconds(PERIOD_W1);
|
|
result[11] = float(MathSin(x != 0 ? 2.0 * M_PI * x : 0));
|
|
x = time / (double)PeriodSeconds(PERIOD_D1);
|
|
result[12] = float(MathSin(x != 0 ? 2.0 * M_PI * x : 0));
|
|
return result;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| UpdateTerminalPath. |
|
|
//+------------------------------------------------------------------+
|
|
void UpdateTerminalPath(const double balance, const double profit, const uint bars,
|
|
double &peak_equity, double &max_drawdown, uint &duration)
|
|
{
|
|
const double equity = balance + profit;
|
|
peak_equity = MathMax(peak_equity, equity);
|
|
max_drawdown = MathMax(max_drawdown, peak_equity - equity);
|
|
duration = MathMax(duration, bars);
|
|
}
|
|
// The historical simulator is evaluated over the complete position path, not a
|
|
// one-bar balance delta. It reports normalized return, drawdown, costs, risk
|
|
// usage and duration; CheckAction below remains its scalar compatibility view.
|
|
bool EvaluateTerminalOutcome(CNeuronRAGMemory &memory, CBufferFloat *action, double balance,
|
|
uint start_position, SRAGTerminalOutcome &outcome)
|
|
{
|
|
const double risk_budget = balance * MemoryRiskBudgetFraction;
|
|
memory.NormalizeTerminalOutcome(0, 0, 0, 0, 0, balance, risk_budget, NForecast, outcome);
|
|
if(!action || start_position >= Rates.Size())
|
|
ReturnFalse;
|
|
double buy_lot = MathMax(double(action[0] - action[3]), 0);
|
|
double sell_lot = MathMax(double(action[3] - action[0]), 0);
|
|
double marg = 0;
|
|
if(!OrderCalcMargin(ORDER_TYPE_BUY, Symb.Name(), 1, Symb.Ask(), marg))
|
|
ReturnFalse;
|
|
double point_cost = Symb.TickValue() / Symb.TickSize();
|
|
if(MathMax(buy_lot, sell_lot) < Symb.LotsMin())
|
|
{
|
|
double loss = -MathMax(Rates[start_position].high - Rates[start_position].open,
|
|
Rates[start_position].open - Rates[start_position].low) *
|
|
point_cost * balance / (2 * marg);
|
|
memory.NormalizeTerminalOutcome(loss, 0, 0, 0, 0, balance, risk_budget, NForecast, outcome);
|
|
return true;
|
|
}
|
|
if((marg * MathMax(buy_lot, sell_lot)) >= balance)
|
|
{
|
|
double loss = -MathMax(Rates[start_position].high - Rates[start_position].open,
|
|
Rates[start_position].open - Rates[start_position].low) *
|
|
point_cost * MathMax(buy_lot, sell_lot);
|
|
memory.NormalizeTerminalOutcome(loss, 0, 0, 0, 0, balance, risk_budget, NForecast, outcome);
|
|
return true;
|
|
}
|
|
point_cost *= MathAbs(buy_lot - sell_lot);
|
|
double tp = 0, sl = 0, profit = 0, reward = 0;
|
|
double peak_equity = balance;
|
|
double max_drawdown = 0;
|
|
double costs = 0;
|
|
double risk_usage = 0;
|
|
uint duration = 0;
|
|
int stops = MathMax(Symb.StopsLevel(), 10);
|
|
int spread = Symb.Spread();
|
|
if(buy_lot > 0)
|
|
{
|
|
tp = action[1] * MaxTP;
|
|
sl = action[2] * MaxSL;
|
|
if(int(tp) < stops || int(sl) < (stops + spread))
|
|
{
|
|
double loss = -MathMax(Rates[start_position].high - Rates[start_position].open,
|
|
Rates[start_position].open - Rates[start_position].low) *
|
|
point_cost * buy_lot;
|
|
memory.NormalizeTerminalOutcome(loss, 0, 0, 0, 0, balance, risk_budget, NForecast, outcome);
|
|
return true;
|
|
}
|
|
tp = (tp + spread) * Symb.Point() + Rates[start_position].open;
|
|
sl = Rates[start_position].open - (sl + spread) * Symb.Point();
|
|
reward = profit = -spread * Symb.Point() * point_cost;
|
|
costs = spread * Symb.Point() * point_cost;
|
|
risk_usage = (Rates[start_position].open - sl) * point_cost;
|
|
for(uint i = start_position; i > 0; i--)
|
|
{
|
|
if(sl >= Rates[i].low)
|
|
{
|
|
double p = (Rates[i].open - sl) * point_cost;
|
|
profit -= p;
|
|
reward -= p * MathPow(DiscFactor, float(i - start_position));
|
|
UpdateTerminalPath(balance, profit, start_position - i + 1, peak_equity, max_drawdown, duration);
|
|
break;
|
|
}
|
|
if(tp <= Rates[i].high)
|
|
{
|
|
double p = (tp - Rates[i].open) * point_cost;
|
|
profit += p;
|
|
reward += p * MathPow(DiscFactor, float(i - start_position));
|
|
UpdateTerminalPath(balance, profit, start_position - i + 1, peak_equity, max_drawdown, duration);
|
|
break;
|
|
}
|
|
double p = (Rates[i - 1].open - Rates[i].open) * point_cost;
|
|
profit += p;
|
|
reward += p * MathPow(DiscFactor, float(i - start_position));
|
|
UpdateTerminalPath(balance, profit, start_position - i + 1, peak_equity, max_drawdown, duration);
|
|
if(-profit >= balance)
|
|
{
|
|
reward -= 1000;
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
if(sell_lot > 0)
|
|
{
|
|
tp = action[4] * MaxTP;
|
|
sl = action[5] * MaxSL;
|
|
if(int(tp) < stops || int(sl) < (stops + spread))
|
|
{
|
|
double loss = -MathMax(Rates[start_position].high - Rates[start_position].open,
|
|
Rates[start_position].open - Rates[start_position].low) *
|
|
point_cost * sell_lot;
|
|
memory.NormalizeTerminalOutcome(loss, 0, 0, 0, 0, balance, risk_budget, NForecast, outcome);
|
|
return true;
|
|
}
|
|
tp = Rates[start_position].open - (tp + spread) * Symb.Point();
|
|
sl = Rates[start_position].open + (sl - spread) * Symb.Point();
|
|
reward = profit = -spread * Symb.Point() * point_cost;
|
|
costs = spread * Symb.Point() * point_cost;
|
|
risk_usage = (sl - Rates[start_position].open) * point_cost;
|
|
for(uint i = start_position; i > 0; i--)
|
|
{
|
|
if(sl <= Rates[i].high)
|
|
{
|
|
double p = (sl - Rates[i].open) * point_cost;
|
|
profit -= p;
|
|
reward -= p * MathPow(DiscFactor, float(i - start_position));
|
|
UpdateTerminalPath(balance, profit, start_position - i + 1, peak_equity, max_drawdown, duration);
|
|
break;
|
|
}
|
|
if(tp >= Rates[i].low)
|
|
{
|
|
double p = (Rates[i].open - tp) * point_cost;
|
|
profit += p;
|
|
reward += p * MathPow(DiscFactor, float(i - start_position));
|
|
UpdateTerminalPath(balance, profit, start_position - i + 1, peak_equity, max_drawdown, duration);
|
|
break;
|
|
}
|
|
double p = (Rates[i - 1].open - Rates[i].open) * point_cost;
|
|
profit -= p;
|
|
reward -= p * MathPow(DiscFactor, float(i - start_position));
|
|
UpdateTerminalPath(balance, profit, start_position - i + 1, peak_equity, max_drawdown, duration);
|
|
if(-profit >= balance)
|
|
{
|
|
reward -= 1000;
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
memory.NormalizeTerminalOutcome(profit, max_drawdown, costs, risk_usage, duration,
|
|
balance, risk_budget, NForecast, outcome);
|
|
return true;
|
|
}
|
|
//+------------------------------------------------------------------+
|
|
//| CheckAction. |
|
|
//+------------------------------------------------------------------+
|
|
double CheckAction(CNeuronRAGMemory &memory, CBufferFloat *action, double balance,
|
|
uint start_position)
|
|
{
|
|
SRAGTerminalOutcome outcome;
|
|
EvaluateTerminalOutcome(memory, action, balance, start_position, outcome);
|
|
return outcome.reward;
|
|
}
|
|
vector<float> OraculAction(const vector<float> &account, CBufferFloat *forecat)
|
|
{
|
|
// Look for target.
|
|
vector<float> result = vector<float>::Zeros(NActions);
|
|
matrix<float> fstate = matrix<float>::Zeros(NForecast, BarDescr);
|
|
if(!forecat.GetData(fstate))
|
|
return result;
|
|
vector<float> target = fstate.Col(0).CumSum();
|
|
if(account[4] > account[5])
|
|
{
|
|
float tp = 0;
|
|
float sl = 0;
|
|
float cur_sl = float(MathMax(MathRand() / 32767.0, 0.01) * MaxSL * Point());
|
|
int pos = 0;
|
|
for(int j = 0; j < NForecast; j++)
|
|
{
|
|
tp = MathMax(tp, target[j] + fstate[j, 1] - fstate[j, 0]);
|
|
pos = j;
|
|
if(cur_sl >= -(target[j] + fstate[j, 2] - fstate[j, 0]))
|
|
break;
|
|
sl = MathMin(sl, target[j] + fstate[j, 2] - fstate[j, 0]);
|
|
}
|
|
if(pos > 0 && tp > 0)
|
|
{
|
|
sl = float(MathMax(MathMin(MathAbs(sl) / (MaxSL * Point()), 1), 0.01));
|
|
tp = float(MathMax(MathMin(tp / (MaxTP * Point()), 1), 0.01));
|
|
result[0] = MathMax(result[0] - result[3], 0.011f);
|
|
result[5] = result[1] = tp;
|
|
result[4] = result[2] = sl;
|
|
result[3] = 0;
|
|
}
|
|
}
|
|
else
|
|
{
|
|
if(account[4] < account[5])
|
|
{
|
|
float tp = 0;
|
|
float sl = 0;
|
|
float cur_sl = float(MathMax(MathRand() / 32767.0, 0.01) * MaxSL * Point());
|
|
int pos = 0;
|
|
for(int j = 0; j < NForecast; j++)
|
|
{
|
|
tp = MathMin(tp, target[j] + fstate[j, 2] - fstate[j, 0]);
|
|
pos = j;
|
|
if(cur_sl <= target[j] + fstate[j, 1] - fstate[j, 0])
|
|
break;
|
|
sl = MathMax(sl, target[j] + fstate[j, 1] - fstate[j, 0]);
|
|
}
|
|
if(pos > 0 && tp < 0)
|
|
{
|
|
sl = float(MathMax(MathMin(MathAbs(sl) / (MaxSL * Point()), 1), 0.01));
|
|
tp = float(MathMax(MathMin(-tp / (MaxTP * Point()), 1), 0.01));
|
|
result[3] = MathMax(result[3] - result[0], 0.011f);
|
|
result[2] = result[4] = tp;
|
|
result[1] = result[5] = sl;
|
|
result[0] = 0;
|
|
}
|
|
}
|
|
else
|
|
{
|
|
ulong argmin = target.ArgMin();
|
|
ulong argmax = target.ArgMax();
|
|
float max_sl = float(MaxSL * Point());
|
|
double equity = account[2] * account[0] * EtalonBalance / (1 + account[1]);
|
|
while(argmax > 0 && argmin > 0)
|
|
{
|
|
if(argmax < argmin && target[argmax] / 2 > MathAbs(target[argmin]) && MathAbs(target[argmin]) < max_sl)
|
|
break;
|
|
if(argmax > argmin && target[argmax] < MathAbs(target[argmin] / 2) && target[argmax] < max_sl)
|
|
break;
|
|
target.Resize(MathMin(argmax, argmin));
|
|
argmin = target.ArgMin();
|
|
argmax = target.ArgMax();
|
|
}
|
|
if(argmin == 0 || (argmax < argmin && argmax > 0))
|
|
{
|
|
float tp = 0;
|
|
float sl = 0;
|
|
float cur_sl = - float(MaxSL * Point());
|
|
ulong pos = 0;
|
|
for(ulong j = 0; j < argmax; j++)
|
|
{
|
|
tp = MathMax(tp, target[j] + fstate[j, 1] - fstate[j, 0]);
|
|
pos = j;
|
|
if(cur_sl >= -(target[j] + fstate[j, 2] - fstate[j, 0]))
|
|
break;
|
|
sl = MathMin(sl, target[j] + fstate[j, 2] - fstate[j, 0]);
|
|
}
|
|
if(pos > 0 && tp > 0)
|
|
{
|
|
sl = (float)MathMax(MathMin(MathAbs(sl) / (MaxSL * Point()), 1), 0.01);
|
|
tp = (float)MathMin(tp / (MaxTP * Point()), 1);
|
|
result[0] = float(MathMax(equity / 100 * 0.01, 0.011));
|
|
result[5] = result[1] = tp;
|
|
result[4] = result[2] = sl;
|
|
result[3] = 0;
|
|
}
|
|
}
|
|
else
|
|
{
|
|
if(argmax == 0 || argmax > argmin)
|
|
{
|
|
float tp = 0;
|
|
float sl = 0;
|
|
float cur_sl = float(MaxSL * Point());
|
|
ulong pos = 0;
|
|
for(ulong j = 0; j < argmin; j++)
|
|
{
|
|
tp = MathMin(tp, target[j] + fstate[j, 2] - fstate[j, 0]);
|
|
pos = j;
|
|
if(cur_sl <= target[j] + fstate[j, 1] - fstate[j, 0])
|
|
break;
|
|
sl = MathMax(sl, target[j] + fstate[j, 1] - fstate[j, 0]);
|
|
}
|
|
if(pos > 0 && tp < 0)
|
|
{
|
|
sl = (float)MathMax(MathMin(MathAbs(sl) / (MaxSL * Point()), 1), 0.01);
|
|
tp = (float)MathMin(-tp / (MaxTP * Point()), 1);
|
|
result[3] = float(MathMax(equity / 100 * 0.01, 0.011));
|
|
result[2] = result[4] = tp;
|
|
result[1] = result[5] = sl;
|
|
result[0] = 0;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
return result;
|
|
}
|
|
|
|
|
|
//| End of Trajectory.mqh |
|