312 lines
14 KiB
MQL5
312 lines
14 KiB
MQL5
|
//+------------------------------------------------------------------+
|
||
|
|
//| StudyOnline.mq5 |
|
||
|
|
//| ORION live direct-transition twin Actor-Critic trainer |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
#property copyright "Copyright DNG®"
|
||
|
|
#property link "https://www.mql5.com/ru/users/dng"
|
||
|
|
#property version "1.00"
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
//| |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
#define StudyOnline
|
||
|
|
#include "Trajectory.mqh"
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
//| |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
input int UpdatePolicy = ActorUpdate;
|
||
|
|
input int UpdateTargets = TargetUpdate;
|
||
|
|
input float Tau = tau;
|
||
|
|
input int CheckpointTransitions = 256;
|
||
|
|
input double MinBalance = 50.0;
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
//| |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
CORIONNet Actor, TargetActor, Q1, Q2, TargetQ1, TargetQ2;
|
||
|
|
CBufferFloat State, TimeState, Account, Action, CriticInput, TargetCriticInput, ScalarTarget;
|
||
|
|
CBufferFloat PreviousAccount;
|
||
|
|
double PreviousBalance = 0, PreviousEquity = 0;
|
||
|
|
double PreviousMarginPenalty = 0;
|
||
|
|
ulong Transition = 0;
|
||
|
|
bool HavePrevious = false;
|
||
|
|
bool PolicyCheckpointPending = false;
|
||
|
|
bool PolicyCheckpointSafe = true;
|
||
|
|
bool ORIONOnlineInitialized = false;
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
//| |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
bool ClearPreviousTransition(void)
|
||
|
|
{
|
||
|
|
HavePrevious = false;
|
||
|
|
PreviousAccount.Clear();
|
||
|
|
PreviousBalance = AccountInfoDouble(ACCOUNT_BALANCE);
|
||
|
|
PreviousEquity = AccountInfoDouble(ACCOUNT_EQUITY);
|
||
|
|
PreviousMarginPenalty = 0;
|
||
|
|
return ORIONMarket.Clear();
|
||
|
|
}
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
//| |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
bool StorePrevious(void)
|
||
|
|
{
|
||
|
|
return (PreviousAccount.AssignArray(GetPointer(Account)) &&
|
||
|
|
(PreviousAccount.GetIndex() < 0 || PreviousAccount.BufferWrite()));
|
||
|
|
}
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
//| Target objects are runtime-only. They start as exact copies of |
|
||
|
|
//| their online counterparts and are never serialized separately. |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
bool LoadOnlineTargets(void)
|
||
|
|
{
|
||
|
|
if(!ORIONLoadPolicyNet(TargetActor, ORION_ACTOR_FILE) ||
|
||
|
|
!ORIONLoadPolicyNet(TargetQ1, ORION_Q1_FILE) || !ORIONLoadPolicyNet(TargetQ2, ORION_Q2_FILE))
|
||
|
|
ReturnFalse;
|
||
|
|
TargetActor.SetOpenCL(ORIONMarket.GetOpenCL());
|
||
|
|
TargetQ1.SetOpenCL(ORIONMarket.GetOpenCL());
|
||
|
|
TargetQ2.SetOpenCL(ORIONMarket.GetOpenCL());
|
||
|
|
if(!ORIONValidatePolicyShape(TargetActor, false) || !ORIONValidatePolicyShape(TargetQ1, true) ||
|
||
|
|
!ORIONValidatePolicyShape(TargetQ2, true))
|
||
|
|
ReturnFalse;
|
||
|
|
TargetActor.TrainMode(false);
|
||
|
|
TargetQ1.TrainMode(false);
|
||
|
|
TargetQ2.TrainMode(false);
|
||
|
|
return true;
|
||
|
|
}
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
//| |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
bool TrainPreviousTransition(const bool terminal)
|
||
|
|
{
|
||
|
|
//--- The live Forecast is at s_next. Online reward is one step, so TD
|
||
|
|
//--- bootstrap uses runtime-only target objects initialized from live files.
|
||
|
|
double q1_target = 0, q2_target = 0;
|
||
|
|
if(!terminal)
|
||
|
|
{
|
||
|
|
if(!TargetActor.feedForward(GetPointer(Account), 1, false, GetPointer(ORIONMarket), -1))
|
||
|
|
ReturnFalse;
|
||
|
|
CNeuronBaseOCL *target_context = TargetActor.Layer(0);
|
||
|
|
CNeuronBaseOCL *target_actor_layer = TargetActor.Layer(3);
|
||
|
|
if(!target_context || !target_actor_layer ||
|
||
|
|
!BuildCriticInput(target_context.getOutput(), target_actor_layer.getOutput(), GetPointer(TargetCriticInput)) ||
|
||
|
|
!TargetQ1.feedForward(GetPointer(TargetCriticInput), 1, false, GetPointer(ORIONMarket), -1) ||
|
||
|
|
!TargetQ2.feedForward(GetPointer(TargetCriticInput), 1, false, GetPointer(ORIONMarket), -1))
|
||
|
|
ReturnFalse;
|
||
|
|
CNeuronBaseOCL *target_q1_layer = TargetQ1.Layer(3), *target_q2_layer = TargetQ2.Layer(3);
|
||
|
|
if(!target_q1_layer || !target_q2_layer)
|
||
|
|
ReturnFalse;
|
||
|
|
//--- TD target is consumed by CPU scalar arithmetic. getResults is the
|
||
|
|
//--- library CPU boundary and also supports a host-only final critic layer.
|
||
|
|
CBufferFloat *qt1 = NULL, *qt2 = NULL;
|
||
|
|
TargetQ1.getResults(qt1);
|
||
|
|
TargetQ2.getResults(qt2);
|
||
|
|
if(!qt1 || !qt2 || qt1.Total() != 1 || qt2.Total() != 1)
|
||
|
|
{
|
||
|
|
DeleteObj(qt1);
|
||
|
|
DeleteObjAndFalse(qt2);
|
||
|
|
}
|
||
|
|
q1_target = double(qt1[0]);
|
||
|
|
q2_target = double(qt2[0]);
|
||
|
|
DeleteObj(qt1);
|
||
|
|
DeleteObj(qt2);
|
||
|
|
}
|
||
|
|
double observed = (PreviousEquity - AccountInfoDouble(ACCOUNT_EQUITY) +
|
||
|
|
PreviousBalance - AccountInfoDouble(ACCOUNT_BALANCE)) * PreviousBalance / EtalonBalance +
|
||
|
|
PreviousMarginPenalty;
|
||
|
|
//--- Match offline ORION: missing a move is penalized at one executable
|
||
|
|
//--- minimum lot, never at a balance-derived virtual position.
|
||
|
|
if(MathMax(double(Account[4]), double(Account[5])) < Symb.LotsMin())
|
||
|
|
{
|
||
|
|
const double point_cost = Symb.TickValue() / Symb.TickSize();
|
||
|
|
if(State.Total() != HistoryBars * BarDescr || Symb.LotsMin() <= 0 ||
|
||
|
|
!MathIsValidNumber(point_cost))
|
||
|
|
ReturnFalse;
|
||
|
|
const double loss = MathAbs(State[0]) * point_cost * Symb.LotsMin();
|
||
|
|
observed -= loss * PreviousBalance / EtalonBalance;
|
||
|
|
}
|
||
|
|
const double target = observed + (terminal ? 0.0 : DiscFactor * MathMin(q1_target, q2_target));
|
||
|
|
if(!MathIsValidNumber(target) || !ScalarTarget.BufferInit(1, 0) || !ScalarTarget.Update(0, float(target)))
|
||
|
|
ReturnFalse;
|
||
|
|
//--- Restore only the ordinary previous state. Forecast scenarios are rerun,
|
||
|
|
//--- never cached or serialized.
|
||
|
|
if(!CreateBuffers(1, GetPointer(State), GetPointer(TimeState), (CBufferFloat*)NULL) ||
|
||
|
|
!ORIONMarket.Clear() ||
|
||
|
|
!ORIONForwardForecastState(GetPointer(State)))
|
||
|
|
ReturnFalse;
|
||
|
|
CNeuronBaseOCL *actor_context = Actor.Layer(0);
|
||
|
|
CNeuronBaseOCL *actor_layer = Actor.Layer(3);
|
||
|
|
if(!actor_context || !actor_layer || !BuildCriticInput(actor_context.getOutput(), actor_layer.getOutput(), GetPointer(CriticInput)) ||
|
||
|
|
!Q1.feedForward(GetPointer(CriticInput), 1, false, GetPointer(ORIONMarket), -1) ||
|
||
|
|
!Q2.feedForward(GetPointer(CriticInput), 1, false, GetPointer(ORIONMarket), -1))
|
||
|
|
ReturnFalse;
|
||
|
|
if(UpdatePolicy > 0 && Transition > 0 && Transition % (ulong)UpdatePolicy == 0 &&
|
||
|
|
!PolicyBackward(Actor, Q1, GetPointer(ScalarTarget), GetPointer(ORIONMarket), -1))
|
||
|
|
ReturnFalse;
|
||
|
|
if(!Q1.backProp(GetPointer(ScalarTarget), GetPointer(ORIONMarket), -1) ||
|
||
|
|
!Q2.backProp(GetPointer(ScalarTarget), GetPointer(ORIONMarket), -1))
|
||
|
|
ReturnFalse;
|
||
|
|
PreviousMarginPenalty = 0;
|
||
|
|
//--- Per-transition guard reads only the three small trainable weight buffers.
|
||
|
|
if(!ORIONVerifyFrozenWeightsExact())
|
||
|
|
ReturnFalse;
|
||
|
|
if(UpdateTargets > 0 && Transition > 0 && Transition % (ulong)UpdateTargets == 0)
|
||
|
|
if(!TargetActor.WeightsUpdate(GetPointer(Actor), Tau) ||
|
||
|
|
!TargetQ1.WeightsUpdate(GetPointer(Q1), Tau) ||
|
||
|
|
!TargetQ2.WeightsUpdate(GetPointer(Q2), Tau))
|
||
|
|
ReturnFalse;
|
||
|
|
return true;
|
||
|
|
}
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
//| |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
int OnInit(void)
|
||
|
|
{
|
||
|
|
ORIONOnlineInitialized = false;
|
||
|
|
if(!ORIONInitIndicators())
|
||
|
|
{ Print("ORION online init: indicators=FAIL"); return INIT_FAILED; }
|
||
|
|
if(!Trade.SetTypeFillingBySymbol(Symb.Name()))
|
||
|
|
{ Print("ORION online init: trade filling=FAIL"); return INIT_FAILED; }
|
||
|
|
if(!ORIONLoadForecastInference())
|
||
|
|
{ Print("ORION online init: forecast=FAIL"); return INIT_FAILED; }
|
||
|
|
if(!ORIONLoadOrCreatePolicySet(Actor, Q1, Q2, false))
|
||
|
|
{ Print("ORION online init: policy checkpoint=FAIL"); return INIT_FAILED; }
|
||
|
|
if(!LoadOnlineTargets())
|
||
|
|
{ Print("ORION online init: runtime targets=FAIL"); return INIT_FAILED; }
|
||
|
|
if(!ORIONVerifyFrozenForecastExact())
|
||
|
|
{ Print("ORION online init: frozen forecast=FAIL"); return INIT_FAILED; }
|
||
|
|
if(!ORIONMarket.Clear())
|
||
|
|
{ Print("ORION online init: forecast clear=FAIL"); return INIT_FAILED; }
|
||
|
|
PreviousBalance = AccountInfoDouble(ACCOUNT_BALANCE);
|
||
|
|
PreviousEquity = AccountInfoDouble(ACCOUNT_EQUITY);
|
||
|
|
ORIONOnlineInitialized = true;
|
||
|
|
return INIT_SUCCEEDED;
|
||
|
|
}
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
//| |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
void OnDeinit(const int reason)
|
||
|
|
{
|
||
|
|
//--- Checkpoint only a fully completed transition. The manifest is written last.
|
||
|
|
if(PolicyCheckpointPending && PolicyCheckpointSafe)
|
||
|
|
if(!ORIONSavePolicySet(Actor, Q1, Q2))
|
||
|
|
PrintFormat("%s -> %d save failed", __FUNCTION__, __LINE__);
|
||
|
|
if(ORIONOnlineInitialized && ORIONForecast != NULL && !ORIONVerifyFrozenForecastExact())
|
||
|
|
PrintFormat("%s -> %d forecast mutation", __FUNCTION__, __LINE__);
|
||
|
|
ORIONForecast = NULL;
|
||
|
|
}
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
//| |
|
||
|
|
//+------------------------------------------------------------------+
|
||
|
|
void OnTick(void)
|
||
|
|
{
|
||
|
|
if(!IsNewBar())
|
||
|
|
return;
|
||
|
|
if(!ORIONRefreshLiveMarket(GetPointer(State), GetPointer(TimeState)))
|
||
|
|
{
|
||
|
|
PrintFormat("%s -> %d market", __FUNCTION__, __LINE__);
|
||
|
|
if(!ClearPreviousTransition())
|
||
|
|
ExpertRemove();
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
//--- Sequential current inference when possible. Training restoration below
|
||
|
|
//--- clears/reruns explicit states, so sequence boundaries remain deliberate.
|
||
|
|
if(!ORIONForwardForecastState(GetPointer(State)))
|
||
|
|
{
|
||
|
|
PrintFormat("%s -> %d forecast", __FUNCTION__, __LINE__);
|
||
|
|
if(!ClearPreviousTransition())
|
||
|
|
ExpertRemove();
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
double buy_value = 0, sell_value = 0;
|
||
|
|
if(!ORIONBuildLiveAccount(PreviousBalance, PreviousEquity, Rates[0].time, GetPointer(Account), buy_value, sell_value))
|
||
|
|
{
|
||
|
|
PrintFormat("%s -> %d account", __FUNCTION__, __LINE__);
|
||
|
|
if(!ClearPreviousTransition())
|
||
|
|
ExpertRemove();
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
//--- A rejected order is recoverable. Loss of the configured minimum
|
||
|
|
//--- deposit is the only financial terminal state for online training.
|
||
|
|
const bool terminal = (AccountInfoDouble(ACCOUNT_BALANCE) <= MinBalance);
|
||
|
|
if(HavePrevious)
|
||
|
|
{
|
||
|
|
PolicyCheckpointSafe = false;
|
||
|
|
if(!TrainPreviousTransition(terminal))
|
||
|
|
{
|
||
|
|
PrintFormat("%s -> %d train: terminal recurrent-state failure", __FUNCTION__, __LINE__);
|
||
|
|
ClearPreviousTransition();
|
||
|
|
ExpertRemove();
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
PolicyCheckpointSafe = true;
|
||
|
|
PolicyCheckpointPending = true;
|
||
|
|
Transition++;
|
||
|
|
if(CheckpointTransitions > 0 && Transition % (ulong)CheckpointTransitions == 0)
|
||
|
|
{
|
||
|
|
if(!ORIONSavePolicySet(Actor, Q1, Q2))
|
||
|
|
{
|
||
|
|
PrintFormat("%s -> %d checkpoint failed", __FUNCTION__, __LINE__);
|
||
|
|
ClearPreviousTransition();
|
||
|
|
ExpertRemove();
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
PolicyCheckpointPending = false;
|
||
|
|
}
|
||
|
|
//--- Training ended at previous state. A terminal transition has no next
|
||
|
|
//--- action, hence it must not rerun the current policy state.
|
||
|
|
if(!terminal && (!CreateBuffers(0, GetPointer(State), GetPointer(TimeState), (CBufferFloat*)NULL) ||
|
||
|
|
!ORIONForwardForecastState(GetPointer(State))))
|
||
|
|
{
|
||
|
|
PrintFormat("%s -> %d restore: terminal recurrent-state failure", __FUNCTION__, __LINE__);
|
||
|
|
ClearPreviousTransition();
|
||
|
|
ExpertRemove();
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
if(terminal)
|
||
|
|
{
|
||
|
|
PrintFormat("ORION online terminal: balance=%.2f min_balance=%.2f", AccountInfoDouble(ACCOUNT_BALANCE), MinBalance);
|
||
|
|
ClearPreviousTransition();
|
||
|
|
ExpertRemove();
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
if(!Actor.feedForward(GetPointer(Account), 1, false, GetPointer(ORIONMarket), -1) ||
|
||
|
|
!ReadAction(Actor, GetPointer(Action)) || !ORIONValidateAction(GetPointer(Action)))
|
||
|
|
{
|
||
|
|
PrintFormat("%s -> %d action", __FUNCTION__, __LINE__);
|
||
|
|
if(!ClearPreviousTransition())
|
||
|
|
ExpertRemove();
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
double margin_penalty = 0;
|
||
|
|
bool market_closed = false;
|
||
|
|
if(!ORIONExecuteAction(GetPointer(Action), buy_value, sell_value, margin_penalty, market_closed))
|
||
|
|
{
|
||
|
|
PrintFormat("%s -> %d execute: terminal partial-trade risk", __FUNCTION__, __LINE__);
|
||
|
|
ClearPreviousTransition();
|
||
|
|
ExpertRemove();
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
if(market_closed)
|
||
|
|
{
|
||
|
|
Print("ORION online execution deferred: market closed");
|
||
|
|
//--- The preceding transition was already trained above. Do not replay it
|
||
|
|
//--- on the first tradable bar, but preserve Actor history for this state.
|
||
|
|
HavePrevious = false;
|
||
|
|
PreviousAccount.Clear();
|
||
|
|
PreviousMarginPenalty = 0;
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
if(margin_penalty < 0)
|
||
|
|
PrintFormat("ORION online insufficient margin: penalty=%.2f", margin_penalty);
|
||
|
|
if(!StorePrevious())
|
||
|
|
{
|
||
|
|
PrintFormat("%s -> %d metadata: terminal post-trade failure", __FUNCTION__, __LINE__);
|
||
|
|
ClearPreviousTransition();
|
||
|
|
ExpertRemove();
|
||
|
|
return;
|
||
|
|
}
|
||
|
|
PreviousBalance = AccountInfoDouble(ACCOUNT_BALANCE);
|
||
|
|
PreviousEquity = AccountInfoDouble(ACCOUNT_EQUITY);
|
||
|
|
PreviousMarginPenalty = margin_penalty;
|
||
|
|
HavePrevious = true;
|
||
|
|
}
|
||
|
|
//+------------------------------------------------------------------+
|