| //+------------------------------------------------------------------+
|
| //| Study.mq5 |
|
| //| Copyright 2026, DNG |
|
| //| https://www.mql5.com/ru/users/dng |
|
| //+------------------------------------------------------------------+
|
| #property copyright "Copyright 2026, DNG"
|
| #property link "https://www.mql5.com/ru/users/dng"
|
| #property version "1.00"
|
| #property strict
|
| #include "Trajectory.mqh"
|
| //---
|
| input group "---- AC-SRM training ----"
|
| input datetime Start = D'2024.01.01'; //Training period start
|
| input datetime End = D'2026.01.01'; //Training period end
|
| input int Iterations = 200000; //Training iterations
|
| input int EpisodeBars = 2 * StackSize; //Bars per episode
|
| input int InpDealHorizon = 24; //Deal evaluation horizon (bars)
|
| input double MinBalance = 50.0; //Minimum account balance
|
| input int UpdatePolicy = 2; //Actor update interval (PolicyUpdatePeriod)
|
| input int InpSeed = 20260921; //Replay seed base (per-owner)
|
| input int InpUpdateTargets = 20; //Target update interval (manifest metadata)
|
| input float InpTauT = 1.0f; //Target copy tau (1.0 = full copy)
|
| //---
|
| input group "---- SRM critic head ----"
|
| input int InpRiskType = defSRM_MEAN_CVAR; //Risk measure (defSRM_*)
|
| input float InpRiskParameter = 0.1f; //Risk parameter (alpha/lambda/nu)
|
| input float InpMeanWeight = 0.0f; //Mean-CVaR omega
|
| input float InpProbWeight = 1.0f; //Distribution probability weight
|
| input float InpValueWeight = 1.0f; //Distribution value weight
|
| input int InpQuantiles = 32; //Head quantiles (factory-fixed)
|
| input int InpSelector = 0; //Policy SRM gradient source (0=owner 1=Q1 2=Q2)
|
| //---
|
| input group "---- AC-SRM production checkpoint ----"
|
| input ENUM_OMPB_STAGE InpOMPBStage = OMPB_STAGE_BASE_POLICY; //Chain stage: keep 03 (base policy)
|
| //---
|
| int Epochs = 0;
|
| CNet Actor;
|
| CNet Q1;
|
| CNet Q2;
|
|
|
| CBufferFloat State;
|
| CBufferFloat TimeState;
|
| CBufferFloat Account;
|
| CBufferFloat NextAccount;
|
| CBufferFloat CurrentAction;
|
| CBufferFloat TeacherAction;
|
| CBufferFloat RandomAction;
|
| CBufferFloat CriticInput;
|
| CBufferFloat SRMValue;
|
| CBufferFloat SRMGrad;
|
|
|
| ulong PolicyTransitions = 0;
|
| ulong PolicyEvents = 0;
|
| ulong PolicySkippedInvalid = 0;
|
| ulong CriticTransitions = 0;
|
| ulong CriticQ1Transitions = 0;
|
| ulong CriticQ2Transitions = 0;
|
| ulong TeacherTransitions = 0;
|
| ulong TeacherCriticTransitions = 0;
|
| ulong RandomCriticTransitions = 0;
|
| ulong CriticQ1Total = 0;
|
| ulong CriticQ2Total = 0;
|
| ulong UnresolvedLabels = 0;
|
| ulong TagTP = 0;
|
| ulong TagSL = 0;
|
| ulong TagHorizon = 0;
|
| double RewardSum = 0;
|
| double CriticQ1Error = 0;
|
| double CriticQ2Error = 0;
|
| bool Stage03StopRequestedLogged = false;
|
| //+-------------------------------------------------------------------+
|
| //| Logs one user-requested Stage 03 stop without misclassifying it. |
|
| //+-------------------------------------------------------------------+
|
| void LogStage03StopRequested(const string phase)
|
| {
|
| if(Stage03StopRequestedLogged)
|
| return;
|
| PrintFormat("ACSRM_STAGE03_STOP_REQUESTED phase=%s", phase);
|
| Stage03StopRequestedLogged = true;
|
| }
|
| //+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
|
| //| Creates the AC-SRM Actor and distribution-Critic architectures. Actor : Base(13)+Cross(3)+Conv(GELU 48->16)+Conv(SIGMOID 16->6). Critic: Base(19)+Cross(5)+Conv(GELU 80->16)+Base(32,None)+ ACSRM head{count=1,window_out=32,None} (factory-fixed). |
|
| //+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
|
| bool CreateActorCriticDescriptions(CArrayObj *&actor_descr, CArrayObj *&critic_descr)
|
| {
|
| actor_descr = new CArrayObj();
|
| critic_descr = new CArrayObj();
|
| if(!actor_descr || !critic_descr)
|
| {
|
| DeleteObj(actor_descr);
|
| DeleteObjAndFalse(critic_descr);
|
| }
|
| actor_descr.FreeMode(true);
|
| critic_descr.FreeMode(true);
|
| //--- Actor: plain scenario trunk without the D2Skill bank layer.
|
| if(!SkillAddBase(actor_descr, AccountDescr) ||
|
| !SkillAddCross(actor_descr, false) ||
|
| !SkillAddConv(actor_descr, 1, (3 * EmbeddingSize), EmbeddingSize, 1, GELU) ||
|
| !SkillAddConv(actor_descr, 1, EmbeddingSize, NActions / 2, 2, SIGMOID))
|
| {
|
| DeleteObj(actor_descr);
|
| DeleteObjAndFalse(critic_descr);
|
| }
|
| //--- Critic: Base(19) feeds the Cross(5) trunk; the last plain Base(32,None)
|
| //--- presents exactly 32 inputs to the factory-fixed ACSRM head.
|
| if(!SkillAddBase(critic_descr, AccountDescr + NActions) ||
|
| !SkillAddCross(critic_descr, true) ||
|
| !SkillAddConv(critic_descr, 1, (5 * EmbeddingSize), EmbeddingSize, 1, GELU) ||
|
| !SkillAddBase(critic_descr, 32))
|
| {
|
| DeleteObj(actor_descr);
|
| DeleteObjAndFalse(critic_descr);
|
| }
|
| CLayerDescription *head = new CLayerDescription();
|
| if(!head)
|
| {
|
| DeleteObj(actor_descr);
|
| DeleteObjAndFalse(critic_descr);
|
| }
|
| head.type = defNeuronACSRM;
|
| head.count = 1;
|
| head.window_out = InpQuantiles;
|
| head.activation = None;
|
| head.optimization = ADAM;
|
| head.batch = BatchSize;
|
| if(!critic_descr.Add(head))
|
| {
|
| DeleteObjAndFalse(head);
|
| DeleteObj(actor_descr);
|
| DeleteObjAndFalse(critic_descr);
|
| }
|
| return(true);
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Validates the AC-SRM policy shapes (no D2 banks, ACSRM heads). |
|
| //+------------------------------------------------------------------+
|
| bool ValidatePolicyShapeACSRM(CNet &actor, CNet &q1, CNet &q2)
|
| {
|
| CNeuronBaseOCL *a0 = actor.Layer(0);
|
| CNeuronBaseOCL *a1 = actor.Layer(1);
|
| CNeuronBaseOCL *a3 = actor.Layer(3);
|
| if(!a0 || !a1 || !a3 || a0.getOutput().Total() != (int)AccountDescr ||
|
| a1.Type() != defNeuronScenarioCrossAttention ||
|
| a3.getOutput().Total() != (int)NActions)
|
| ReturnFalse;
|
| for(int i = 0; i < 2; i++)
|
| {
|
| CNet *critic = (i == 0 ? GetPointer(q1) : GetPointer(q2));
|
| CNeuronBaseOCL *c0 = critic.Layer(0);
|
| CNeuronBaseOCL *c3 = critic.Layer(3);
|
| CNeuronBaseOCL *c4 = critic.Layer(4);
|
| if(!c0 || !c3 || !c4 ||
|
| c0.getOutput().Total() != (int)(AccountDescr + NActions) ||
|
| c3.getOutput().Total() != 32 ||
|
| c4.Type() != defNeuronACSRM ||
|
| c4.getOutput().Total() != 1)
|
| ReturnFalse;
|
| }
|
| return(true);
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Creates fresh AC-SRM Actor + Q1 + Q2 on the Market context. |
|
| //+------------------------------------------------------------------+
|
| bool CreatePolicySet(void)
|
| {
|
| CArrayObj *actor_descr = NULL, *critic_descr = NULL;
|
| if(!CreateActorCriticDescriptions(actor_descr, critic_descr))
|
| {
|
| DeleteObj(actor_descr);
|
| DeleteObj(critic_descr);
|
| ReturnFalse;
|
| }
|
| bool result = Actor.Create(actor_descr);
|
| if(result)
|
| result = Actor.SetOpenCLChecked(SkillMarket.GetOpenCL());
|
| if(result)
|
| result = Q1.Create(critic_descr);
|
| if(result)
|
| result = Q1.SetOpenCLChecked(SkillMarket.GetOpenCL());
|
| if(result)
|
| result = Q2.Create(critic_descr);
|
| if(result)
|
| result = Q2.SetOpenCLChecked(SkillMarket.GetOpenCL());
|
| DeleteObj(actor_descr);
|
| DeleteObj(critic_descr);
|
| return(result);
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Loads or creates ACSRMActor.nnw / ACSRMQ1.nnw / ACSRMQ2.nnw. |
|
| //+------------------------------------------------------------------+
|
| bool LoadOrCreatePolicies(void)
|
| {
|
| if(!SkillACRecoverTransaction(Skill_ACTOR_FILE, Skill_Q1_FILE, Skill_Q2_FILE,
|
| Skill_AC_MANIFEST_FILE, Skill_ACTOR_NEXT_FILE,
|
| Skill_Q1_NEXT_FILE, Skill_Q2_NEXT_FILE,
|
| Skill_AC_MANIFEST_NEXT_FILE, Skill_ACTOR_PREVIOUS_FILE,
|
| Skill_Q1_PREVIOUS_FILE, Skill_Q2_PREVIOUS_FILE,
|
| Skill_AC_MANIFEST_PREVIOUS_FILE,
|
| Skill_AC_TRANSACTION_FILE))
|
| {
|
| Print("Skill policy recovery=FAIL; incomplete checkpoint requires repair");
|
| ReturnFalse;
|
| }
|
| const bool manifest_exists = FileIsExist(Skill_AC_MANIFEST_FILE, FILE_COMMON);
|
| bool loaded = false;
|
| if(manifest_exists)
|
| {
|
| loaded = (SkillValidateACManifestFile(true, Skill_AC_MANIFEST_FILE, false) &&
|
| SkillLoadPolicyNet(Actor, Skill_ACTOR_FILE) &&
|
| SkillLoadPolicyNet(Q1, Skill_Q1_FILE) &&
|
| SkillLoadPolicyNet(Q2, Skill_Q2_FILE));
|
| if(!loaded)
|
| Print("Skill policy restore=FAIL; recreating incompatible policy set");
|
| }
|
| if(!loaded)
|
| {
|
| if(!CreatePolicySet())
|
| ReturnFalse;
|
| }
|
| if(!SkillBindPolicyOpenCLChecked(Actor, Q1, Q2) ||
|
| !ValidatePolicyShapeACSRM(Actor, Q1, Q2))
|
| ReturnFalse;
|
| Actor.TrainMode(true);
|
| Q1.TrainMode(true);
|
| Q2.TrainMode(true);
|
| return(true);
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Saves the Stage 03 AC-SRM policy checkpoint. |
|
| //+------------------------------------------------------------------+
|
| bool SavePolicies(void)
|
| {
|
| return(SkillSaveStage03StopCheckpoint(Actor, Q1, Q2));
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Implements PrepareHistory. |
|
| //+------------------------------------------------------------------+
|
| bool PrepareHistory(int &first_position, int &last_position)
|
| {
|
| int start = iBarShift(Symb.Name(), TimeFrame, Start);
|
| int end = iBarShift(Symb.Name(), TimeFrame, End);
|
| int bars = CopyRates(Symb.Name(), TimeFrame, 0, start, Rates);
|
| if(bars <= 0 || !RSI.BufferResize(bars) || !CCI.BufferResize(bars) ||
|
| !ATR.BufferResize(bars) || !MACD.BufferResize(bars))
|
| ReturnFalse;
|
| int wait = -1;
|
| bool calculated = false;
|
| do
|
| {
|
| calculated = (RSI.BarsCalculated() >= bars && CCI.BarsCalculated() >= bars &&
|
| ATR.BarsCalculated() >= bars && MACD.BarsCalculated() >= bars);
|
| Sleep(100);
|
| wait++;
|
| }
|
| while(!calculated && wait < 100);
|
| if(!calculated)
|
| ReturnFalse;
|
| RSI.Refresh();
|
| CCI.Refresh();
|
| ATR.Refresh();
|
| MACD.Refresh();
|
| if(!ArraySetAsSeries(Rates, true))
|
| ReturnFalse;
|
| //--- The deal outcome must be resolvable INSIDE the training window: the
|
| //--- full horizon is reserved past the End boundary so every label
|
| //--- (TP/SL/HORIZON) completes within [Start, End].
|
| first_position = end + MathMax(1, SkillManifestDealHorizon);
|
| //--- CreateBuffers(position,...,forecast) forms its state from
|
| //--- position+NForecast and needs HistoryBars bars behind it.
|
| last_position = start - HistoryBars - NForecast;
|
| return(last_position > first_position);
|
| }
|
| //+-------------------------------------------------------------------+
|
| //| Applies the distribution target (realized return) to one critic. |
|
| //+-------------------------------------------------------------------+
|
| bool UpdateCriticDistribution(CNet &critic, const double reward, const bool is_q1)
|
| {
|
| CNeuronBaseOCL *base = critic.Layer(3);
|
| CNeuronBaseOCL *head = critic.Layer(4);
|
| if(!base || !head || head.Type() != defNeuronACSRM || !MathIsValidNumber(reward) ||
|
| base.getOutput().Total() != 32)
|
| ReturnFalse;
|
| CNeuronACSRM *acsrm = (CNeuronACSRM*)head;
|
| if(!acsrm.calcDistributionTargetGradients(base, float(reward),
|
| InpProbWeight, InpValueWeight))
|
| ReturnFalse;
|
| //--- Real ACSRM critic error: RMSE of the FQF centroids from the realized
|
| //--- outcome. backPropGradient never updates getRecentAverageError(), so
|
| //--- the old dashboard rRMSE stayed 0 forever; track the loss here.
|
| double rmse = -1;
|
| {
|
| CBufferFloat *quantiles = NULL, *probabilities = NULL;
|
| if(acsrm.GetDistributionBuffers(quantiles, probabilities) &&
|
| quantiles != NULL && quantiles.Total() > 0 && quantiles.BufferRead())
|
| {
|
| double sum_sq = 0;
|
| bool valid = true;
|
| for(int i = 0; i < quantiles.Total(); i++)
|
| {
|
| const double z = double(quantiles[i]);
|
| if(!MathIsValidNumber((float)z))
|
| { valid = false; break; }
|
| sum_sq += (z - reward) * (z - reward);
|
| }
|
| if(valid)
|
| rmse = MathSqrt(sum_sq / double(quantiles.Total()));
|
| }
|
| }
|
| if(rmse >= 0)
|
| {
|
| if(is_q1)
|
| CriticQ1Error += (rmse - CriticQ1Error) * 0.01;
|
| else
|
| CriticQ2Error += (rmse - CriticQ2Error) * 0.01;
|
| }
|
| return(critic.backPropGradient(GetPointer(SkillMarket), -1, -1, true));
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Actor policy update through the frozen critic's SRM. |
|
| //+------------------------------------------------------------------+
|
| int SelectPolicyCritic(const int event_index)
|
| {
|
| //--- Exactly ONE Critic is the gradient source per policy-update event.
|
| //--- The default selector is balanced over Actor-update EVENTS (a hash of
|
| //--- the event counter and the replay seed), NOT over the Owner of the
|
| //--- training example: with a sequential position-- sampler consecutive
|
| //--- policy events inside a stretch would share one Owner parity.
|
| switch(InpSelector)
|
| {
|
| case 1: return(0);
|
| case 2: return(1);
|
| default: return(int((ulong(MathMax(event_index, 0)) * ulong(2654435761u) +
|
| ulong(InpSeed)) & 1));
|
| }
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Returns the scalar SRM value of one Critic head distribution. |
|
| //+------------------------------------------------------------------+
|
| bool CriticSRMValue(CNet &critic, float &value)
|
| {
|
| CNeuronBaseOCL *head = critic.Layer(4);
|
| if(!head || head.Type() != defNeuronACSRM)
|
| ReturnFalse;
|
| CNeuronACSRM *acsrm = (CNeuronACSRM*)head;
|
| if(!SRMValue.BufferInit(1, 0) || !SRMValue.BufferCreate(SkillMarket.GetOpenCL()) ||
|
| !SRMValue.BufferWrite())
|
| ReturnFalse;
|
| if(!acsrm.SRMForward(GetPointer(SRMValue), 1, InpRiskType, InpRiskParameter, InpMeanWeight))
|
| ReturnFalse;
|
| if(!SRMValue.BufferRead())
|
| ReturnFalse;
|
| value = float(SRMValue[0]);
|
| return(MathIsValidNumber(value));
|
| }
|
| //+----------------------------------------------------------------------------------------------------------------+
|
| //| Reports whether the last SRM refusal was a distribution-contract rejection rather than an execution failure. |
|
| //+----------------------------------------------------------------------------------------------------------------+
|
| bool CriticLastSRMInvalid(CNet &critic)
|
| {
|
| CNeuronBaseOCL *head = critic.Layer(4);
|
| if(!head || head.Type() != defNeuronACSRM)
|
| return(false);
|
| return(((CNeuronACSRM*)head).LastSRMInvalid());
|
| }
|
| //+-----------------------------------------------------------------------------------+
|
| //| Actor policy update through the frozen critic's SRM (PolicyOutputGradientACSRM). |
|
| //+-----------------------------------------------------------------------------------+
|
| bool PolicyOutputGradientACSRM(CNet &actor, CNet &critic, const float upstream = 1.0f)
|
| {
|
| CNeuronBaseOCL *critic_head = critic.Layer(4);
|
| CNeuronBaseOCL *critic_base = critic.Layer(0);
|
| CNeuronBaseOCL *actor_output = actor.Layer(3);
|
| if(!critic_head || critic_head.Type() != defNeuronACSRM || !critic_base || !actor_output ||
|
| critic_base.getGradient().Total() != (int)(AccountDescr + NActions) ||
|
| actor_output.getGradient().Total() != (int)NActions)
|
| ReturnFalse;
|
| CNeuronACSRM *acsrm = (CNeuronACSRM*)critic_head;
|
| if(!SRMValue.BufferInit(1, 0) || !SRMValue.BufferCreate(SkillMarket.GetOpenCL()) ||
|
| !SRMValue.BufferWrite())
|
| ReturnFalse;
|
| if(!acsrm.SRMForward(GetPointer(SRMValue), 1, InpRiskType, InpRiskParameter, InpMeanWeight))
|
| ReturnFalse;
|
| if(!SRMGrad.BufferInit(1, upstream) || !SRMGrad.BufferCreate(SkillMarket.GetOpenCL()) ||
|
| !SRMGrad.BufferWrite())
|
| ReturnFalse;
|
| if(!acsrm.SRMGradient(GetPointer(SRMGrad), 1, InpRiskType, InpRiskParameter, InpMeanWeight))
|
| ReturnFalse;
|
| //--- Freeze critic weights while propagating the policy gradient.
|
| if(!critic.SetWeightsUpdate(false))
|
| ReturnFalse;
|
| bool result = critic.backPropGradient(GetPointer(SkillMarket), -1, -1, true);
|
| if(!critic.SetWeightsUpdate(true))
|
| result = false;
|
| CBufferFloat account_gradient;
|
| if(!(result &&
|
| account_gradient.BufferInit(AccountDescr, 0) &&
|
| (account_gradient.GetIndex() >= 0 || account_gradient.BufferCreate(SkillMarket.GetOpenCL())) &&
|
| SkillDevice.Split2(GetPointer(account_gradient), actor_output.getGradient(),
|
| critic_base.getGradient(), AccountDescr, NActions) &&
|
| ApplyActorSigmoidDerivative(actor_output)))
|
| result = false;
|
| return(result);
|
| }
|
| //+---------------------------------------------------------------------------------------------------------------------+
|
| //| Accumulates one supervised signal (teacher/random) into the actor output-layer gradient: (t - a) .* a .* (1 - a). |
|
| //+---------------------------------------------------------------------------------------------------------------------+
|
| bool AddActorSupervisedGradient(CNet &actor, CBufferFloat *target)
|
| {
|
| CNeuronBaseOCL *actor_output = actor.Layer(3);
|
| if(!actor_output || !target || target.Total() != (int)NActions ||
|
| actor_output.getOutput().Total() != (int)NActions ||
|
| actor_output.getGradient().Total() != (int)NActions || target.GetIndex() < 0 ||
|
| actor_output.getOutput().GetIndex() < 0 || actor_output.getGradient().GetIndex() < 0)
|
| ReturnFalse;
|
| if(!ActorDerivOneMinus.BufferInit(NActions, 0) ||
|
| !ActorDerivOneMinus.BufferCreate(SkillMarket.GetOpenCL()) ||
|
| !ActorDerivOneMinus.BufferWrite() ||
|
| !ActorGradScratch.BufferInit(NActions, 0) ||
|
| !ActorGradScratch.BufferCreate(SkillMarket.GetOpenCL()) ||
|
| !ActorGradScratch.BufferWrite() ||
|
| !SkillDevice.Bind(SkillMarket.GetOpenCL()))
|
| ReturnFalse;
|
| //--- diff = t - a; scaled = diff * a * (1 - a); accumulate on device.
|
| //--- This reproduces the library's calcOutputGradients for a SIGMOID
|
| //--- output exactly, but pre-update and summed before ONE weight update.
|
| if(!SkillDevice.Subtract(target, actor_output.getOutput(), GetPointer(ActorDerivOneMinus), NActions) ||
|
| !SkillDevice.DeActivationOnDevice(actor_output.getOutput(), GetPointer(ActorGradScratch),
|
| GetPointer(ActorDerivOneMinus), SIGMOID) ||
|
| !SkillDevice.Add(actor_output.getGradient(), GetPointer(ActorGradScratch),
|
| actor_output.getGradient(), NActions))
|
| ReturnFalse;
|
| return(true);
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Implements TrainTransition. |
|
| //+------------------------------------------------------------------+
|
| bool TrainTransition(const int position, const int iteration, bool &terminal)
|
| {
|
| terminal = false;
|
| const bool trace_td = (iteration < 24 || iteration % 10000 == 0 || iteration >= Iterations - 1);
|
| if(!SkillForwardForecast(position, GetPointer(State), GetPointer(TimeState)))
|
| { Print("TrainTransition stage=current_forecast"); ReturnFalse; }
|
| double reward = 0;
|
| if(!Actor.feedForward(GetPointer(Account), 1, false, GetPointer(SkillMarket), -1) ||
|
| !ReadAction(Actor, GetPointer(CurrentAction)) ||
|
| !AdvanceAccount(GetPointer(Account), GetPointer(CurrentAction), position, MinBalance,
|
| GetPointer(NextAccount), reward, terminal))
|
| { Print("TrainTransition stage=account_transition"); ReturnFalse; }
|
| const double buy_lot = MathMax(0.0, double(CurrentAction[0] - CurrentAction[3]));
|
| const double sell_lot = MathMax(0.0, double(CurrentAction[3] - CurrentAction[0]));
|
| if(trace_td)
|
| PrintFormat("ACSRM action iteration=%d balance=%.2f buy_lot=%.8f buy_tp=%.8f buy_sl=%.8f " +
|
| "sell_lot=%.8f sell_tp=%.8f sell_sl=%.8f",
|
| iteration, MathMax(0.0, double(Account[0]) * EtalonBalance),
|
| buy_lot, double(CurrentAction[1]), double(CurrentAction[2]),
|
| sell_lot, double(CurrentAction[4]), double(CurrentAction[5]));
|
| if(trace_td)
|
| PrintFormat("ACSRM return iteration=%d reward=%.8f balance=%.2f terminal=%s",
|
| iteration, reward, MathMax(0.0, double(NextAccount[0]) * EtalonBalance),
|
| (terminal ? "true" : "false"));
|
| //--- Critic input is built from the live device activations.
|
| CNeuronBaseOCL *context_layer = Actor.Layer(0);
|
| CNeuronBaseOCL *actor_layer = Actor.Layer(3);
|
| if(!context_layer || !actor_layer ||
|
| !BuildCriticInput(context_layer.getOutput(), actor_layer.getOutput(), GetPointer(CriticInput)))
|
| { Print("TrainTransition stage=critic_input"); ReturnFalse; }
|
| if(!Q1.feedForward(GetPointer(CriticInput), 1, false, GetPointer(SkillMarket), -1) ||
|
| !Q2.feedForward(GetPointer(CriticInput), 1, false, GetPointer(SkillMarket), -1))
|
| { Print("TrainTransition stage=critic_forward"); ReturnFalse; }
|
| //--- Offline target: the COMPLETED result of the specific deal fixed at this
|
| //--- bar over the configured horizon (TP first / SL first / HORIZON outcome),
|
| //--- discounted by elapsed bars. Realized future bars may label the deal but
|
| //--- never enter the decision state (SkillForwardForecast already shifts the
|
| //--- window to the strictly-past bars). It is a finished result and needs no
|
| //--- target bootstrap.
|
| int deal_tag = 0;
|
| const double deal_y = EvaluateAction(GetPointer(CurrentAction),
|
| MathMax(0.0, double(Account[0]) * EtalonBalance),
|
| (uint)position, InpDealHorizon, deal_tag);
|
| if(!MathIsValidNumber(deal_y))
|
| { Print("TrainTransition stage=deal_label"); ReturnFalse; }
|
| if(trace_td)
|
| PrintFormat("ACSRM offline iteration=%d outcome=%.8f tag=%d advance_reward=%.8f horizon=%d",
|
| iteration, deal_y, deal_tag, reward, InpDealHorizon);
|
| //--- Offline Owner separation: the example id (position,slot,seed) fixes one
|
| //--- stable Owner-Critic; re-reading the same record keeps its label there.
|
| const int owner = SkillObservationOwner(position, 0, InpSeed);
|
| if(deal_tag == 0)
|
| UnresolvedLabels++;
|
| else
|
| {
|
| CNet *owner_critic = (owner == 0 ? GetPointer(Q1) : GetPointer(Q2));
|
| if(!owner_critic.feedForward(GetPointer(CriticInput), 1, false, GetPointer(SkillMarket), -1) ||
|
| !UpdateCriticDistribution(owner_critic, deal_y, (owner_critic == GetPointer(Q1))))
|
| {
|
| if(IsStopped())
|
| {
|
| LogStage03StopRequested("critic_distribution_backward");
|
| return(false);
|
| }
|
| Print("TrainTransition stage=critic_distribution_backward");
|
| ReturnFalse;
|
| }
|
| CriticTransitions++;
|
| if(owner_critic == GetPointer(Q1))
|
| {
|
| CriticQ1Transitions++;
|
| CriticQ1Total++;
|
| }
|
| else
|
| {
|
| CriticQ2Transitions++;
|
| CriticQ2Total++;
|
| }
|
| RewardSum += deal_y;
|
| if(deal_tag == 1)
|
| TagTP++;
|
| else
|
| if(deal_tag == 2)
|
| TagSL++;
|
| else
|
| TagHorizon++;
|
| }
|
| //--- Actor multi-objective update. Policy, teacher and random signals are
|
| //--- computed on the SAME forward activations and the SAME pre-update
|
| //--- weights and ACCUMULATED into the output-layer gradient; ONE
|
| //--- backPropGradient applies them together. No weight change happens
|
| //--- between the contributions so credit assignment stays consistent.
|
| bool actor_update_needed = false;
|
| bool actor_update_blocked = false;
|
| //--- Gradient hygiene: clear the Actor output gradient once BEFORE the
|
| //--- current observation accumulates its contributions. A deferred policy
|
| //--- event (UpdatePolicy > 1) would otherwise Add teacher/random on top of
|
| //--- the stale gradient the previous combined update consumed, since Split2
|
| //--- only overwrites the gradient on the policy event.
|
| if(actor_layer.getGradient().Total() != (int)NActions ||
|
| !actor_layer.getGradient().Fill(0))
|
| { Print("TrainTransition stage=actor_gradient_reset"); ReturnFalse; }
|
| //--- Policy gate: a RESOLVABLE deal label (tag != 0). Requiring broker
|
| //--- executability here would give the cold-start Actor no gradient while
|
| //--- its lots are below LotsMin(); with the min-lot floor inside
|
| //--- CheckAction the gradient direction exists from the first iteration.
|
| const bool resolvable_action = (deal_tag != 0);
|
| if(iteration > 0 && UpdatePolicy > 0 && iteration % UpdatePolicy == 0)
|
| {
|
| if(resolvable_action)
|
| {
|
| //--- Fresh Critic activations on the same observation AFTER the Critic
|
| //--- weight update: the policy gradient must not mix pre-update
|
| //--- activations with post-update weights.
|
| if(!Q1.feedForward(GetPointer(CriticInput), 1, false, GetPointer(SkillMarket), -1) ||
|
| !Q2.feedForward(GetPointer(CriticInput), 1, false, GetPointer(SkillMarket), -1))
|
| { Print("TrainTransition stage=policy_critic_forward"); ReturnFalse; }
|
| const int pick = SelectPolicyCritic(int(PolicyEvents++));
|
| CNet *policy_critic = (pick == 0 ? GetPointer(Q1) : GetPointer(Q2));
|
| float policy_srm = 0.0f;
|
| if(!CriticSRMValue(*policy_critic, policy_srm))
|
| {
|
| if(CriticLastSRMInvalid(*policy_critic))
|
| {
|
| PolicySkippedInvalid++;
|
| actor_update_blocked = true;
|
| if(trace_td)
|
| PrintFormat("ACSRM policy skipped iteration=%d reason=invalid_srm_distribution", iteration);
|
| }
|
| else
|
| { Print("TrainTransition stage=policy_critic_srm"); ReturnFalse; }
|
| }
|
| else
|
| {
|
| const bool policy_ok = PolicyOutputGradientACSRM(Actor, *policy_critic, 1.0f);
|
| if(!policy_ok)
|
| {
|
| if(CriticLastSRMInvalid(*policy_critic))
|
| {
|
| PolicySkippedInvalid++;
|
| actor_update_blocked = true;
|
| if(trace_td)
|
| PrintFormat("ACSRM policy skipped iteration=%d reason=invalid_srm_gradient", iteration);
|
| }
|
| else
|
| { Print("TrainTransition stage=policy_backward"); ReturnFalse; }
|
| }
|
| else
|
| {
|
| PolicyTransitions++;
|
| actor_update_needed = true;
|
| }
|
| }
|
| }
|
| else
|
| PolicySkippedInvalid++;
|
| }
|
| //--- Restored original-project Actor signals: a profitable realized-future
|
| //--- teacher and a profitable random executable action re-enter the action
|
| //--- manifold; their deals also expand Owner-Critic coverage. Only
|
| //--- RESOLVABLE labels (tag != 0) train anything: with the SL sign fixed
|
| //--- the reward > 0 gate now stands for a genuine TP profit.
|
| double teacher_reward = 0;
|
| int teacher_tag = 0;
|
| if(!BuildTeacherAction(position, GetPointer(Account), GetPointer(TeacherAction),
|
| teacher_reward, InpDealHorizon, teacher_tag))
|
| { Print("TrainTransition stage=teacher_action"); ReturnFalse; }
|
| if(trace_td)
|
| PrintFormat("ACSRM teacher iteration=%d reward=%.8f tag=%d", iteration, teacher_reward, teacher_tag);
|
| if(teacher_reward > 0.0 && teacher_tag != 0)
|
| {
|
| if(!AddActorSupervisedGradient(Actor, GetPointer(TeacherAction)))
|
| {
|
| if(IsStopped())
|
| {
|
| LogStage03StopRequested("teacher_actor_backward");
|
| return(false);
|
| }
|
| Print("TrainTransition stage=teacher_actor_backward");
|
| ReturnFalse;
|
| }
|
| TeacherTransitions++;
|
| actor_update_needed = true;
|
| }
|
| if(teacher_tag == 0)
|
| UnresolvedLabels++;
|
| else
|
| {
|
| const int teacher_owner = SkillObservationOwner(position, 1, InpSeed);
|
| CNet *teacher_critic = (teacher_owner == 0 ? GetPointer(Q1) : GetPointer(Q2));
|
| if(!BuildCriticInput(context_layer.getOutput(), GetPointer(TeacherAction), GetPointer(CriticInput)) ||
|
| !teacher_critic.feedForward(GetPointer(CriticInput), 1, false, GetPointer(SkillMarket), -1) ||
|
| !UpdateCriticDistribution(teacher_critic, teacher_reward, (teacher_critic == GetPointer(Q1))))
|
| {
|
| if(IsStopped())
|
| {
|
| LogStage03StopRequested("teacher_critic_backward");
|
| return(false);
|
| }
|
| Print("TrainTransition stage=teacher_critic_backward");
|
| ReturnFalse;
|
| }
|
| TeacherCriticTransitions++;
|
| if(teacher_critic == GetPointer(Q1))
|
| CriticQ1Total++;
|
| else
|
| CriticQ2Total++;
|
| }
|
| double random_reward = 0;
|
| int random_tag = 0;
|
| if(!BuildRandomAction(position, GetPointer(Account), GetPointer(RandomAction),
|
| random_reward, InpDealHorizon, random_tag))
|
| { Print("TrainTransition stage=random_action"); ReturnFalse; }
|
| if(trace_td)
|
| PrintFormat("ACSRM random iteration=%d reward=%.8f tag=%d", iteration, random_reward, random_tag);
|
| if(random_reward > 0.0 && random_tag != 0)
|
| {
|
| if(!AddActorSupervisedGradient(Actor, GetPointer(RandomAction)))
|
| {
|
| if(IsStopped())
|
| {
|
| LogStage03StopRequested("random_actor_backward");
|
| return(false);
|
| }
|
| Print("TrainTransition stage=random_actor_backward");
|
| ReturnFalse;
|
| }
|
| actor_update_needed = true;
|
| }
|
| if(random_tag == 0)
|
| UnresolvedLabels++;
|
| else
|
| {
|
| const int random_owner = SkillObservationOwner(position, 2, InpSeed);
|
| CNet *random_critic = (random_owner == 0 ? GetPointer(Q1) : GetPointer(Q2));
|
| if(!BuildCriticInput(context_layer.getOutput(), GetPointer(RandomAction), GetPointer(CriticInput)) ||
|
| !random_critic.feedForward(GetPointer(CriticInput), 1, false, GetPointer(SkillMarket), -1) ||
|
| !UpdateCriticDistribution(random_critic, random_reward, (random_critic == GetPointer(Q1))))
|
| {
|
| if(IsStopped())
|
| {
|
| LogStage03StopRequested("random_critic_backward");
|
| return(false);
|
| }
|
| Print("TrainTransition stage=random_critic_backward");
|
| ReturnFalse;
|
| }
|
| RandomCriticTransitions++;
|
| if(random_critic == GetPointer(Q1))
|
| CriticQ1Total++;
|
| else
|
| CriticQ2Total++;
|
| }
|
| //--- ONE coherent Actor weight update with the combined gradient (if any).
|
| //--- An invalid SRM distribution blocks the whole Actor update for this
|
| //--- observation, matching srm_invalid_policy=skip_actor_update.
|
| if(actor_update_needed && !actor_update_blocked &&
|
| !Actor.backPropGradient(GetPointer(SkillMarket), -1, -1, true))
|
| {
|
| if(IsStopped())
|
| {
|
| LogStage03StopRequested("actor_combined_backward");
|
| return(false);
|
| }
|
| Print("TrainTransition stage=actor_combined_backward");
|
| ReturnFalse;
|
| }
|
| return(true);
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Implements TrainACSRMActorCritic. |
|
| //+------------------------------------------------------------------+
|
| void TrainACSRMActorCritic(void)
|
| {
|
| Stage03StopRequestedLogged = false;
|
| int first = 0, last = 0;
|
| if(!PrepareHistory(first, last))
|
| { PrintFormat("%s -> %d history unavailable", __FUNCTION__, __LINE__); return; }
|
| uint shown = GetTickCount();
|
| int completed_iterations = 0;
|
| int failed_iteration = -1;
|
| int failed_position = -1;
|
| int position = last;
|
| int episode = 0;
|
| PolicyTransitions = 0;
|
| PolicyEvents = 0;
|
| PolicySkippedInvalid = 0;
|
| CriticTransitions = 0;
|
| CriticQ1Transitions = 0;
|
| CriticQ2Transitions = 0;
|
| TeacherTransitions = 0;
|
| TeacherCriticTransitions = 0;
|
| RandomCriticTransitions = 0;
|
| CriticQ1Total = 0;
|
| CriticQ2Total = 0;
|
| UnresolvedLabels = 0;
|
| TagTP = 0;
|
| TagSL = 0;
|
| TagHorizon = 0;
|
| RewardSum = 0;
|
| CriticQ1Error = 0;
|
| CriticQ2Error = 0;
|
| for(int iteration = 0; iteration < Iterations && !IsStopped(); iteration++)
|
| {
|
| //--- Fresh episode: seed the sampler from the observation Owner.
|
| if(episode == 0)
|
| {
|
| if(!SkillForwardForecast(position, GetPointer(State), GetPointer(TimeState)))
|
| break;
|
| MathSrand(InpSeed + position);
|
| const vector<float> sampled = SampleAccount(GetPointer(State), Rates[position].time,
|
| EtalonBalance, MinBalance);
|
| if(sampled.Size() != AccountDescr || !Account.AssignArray(sampled) ||
|
| (Account.GetIndex() < 0 && !Account.BufferCreate(SkillMarket.GetOpenCL())) ||
|
| !Account.BufferWrite() || !Actor.Clear())
|
| break;
|
| }
|
| bool terminal = false;
|
| if(!TrainTransition(position, iteration, terminal))
|
| {
|
| if(IsStopped())
|
| {
|
| LogStage03StopRequested("train_transition");
|
| break;
|
| }
|
| failed_iteration = iteration;
|
| failed_position = position;
|
| break;
|
| }
|
| completed_iterations++;
|
| if((NextAccount.GetIndex() >= 0 && !NextAccount.BufferRead()) ||
|
| !Account.AssignArray(GetPointer(NextAccount)) ||
|
| (Account.GetIndex() >= 0 && !Account.BufferWrite()))
|
| {
|
| failed_iteration = iteration;
|
| failed_position = position;
|
| break;
|
| }
|
| position--;
|
| episode++;
|
| //--- Close the episode on range exhaustion, its bar limit or a
|
| //--- terminal account state.
|
| if(position < first || episode >= MathMax(1, EpisodeBars) || terminal)
|
| {
|
| if(iteration % 10000 == 0)
|
| PrintFormat("ACSRM episode closed iteration=%d episode=%d position=%d terminal=%s",
|
| iteration, episode, position, (terminal ? "true" : "false"));
|
| if(position < first)
|
| position = last;
|
| episode = 0;
|
| }
|
| //--- Refresh the progress panel without delaying training updates.
|
| if(GetTickCount() - shown > 500)
|
| {
|
|
|
|
|
|
|
|
|
| const double r_mean = (CriticTransitions > 0 ? RewardSum / double(CriticTransitions) : 0.0);
|
| Comment(StringFormat("ACSRM AC %6.2f%% Q1 err %.8f Q2 err %.8f critic %I64u " +
|
| "policy %I64u skip %I64u Rmean %.4f",
|
| 100.0 * iteration / MathMax(Iterations, 1), CriticQ1Error, CriticQ2Error,
|
| CriticTransitions, PolicyTransitions, PolicySkippedInvalid,
|
| r_mean));
|
| shown = GetTickCount();
|
| }
|
| }
|
| if(IsStopped())
|
| {
|
| //--- A stop can land between parameter updates; never publish a partial step.
|
| LogStage03StopRequested("training_loop");
|
| Comment("");
|
| PrintFormat("ACSRM_STAGE03_STOP_NO_PUBLISH completed=%d/%d critic=%I64u q1=%I64u q2=%I64u policy=%I64u",
|
| completed_iterations, Iterations, CriticTransitions,
|
| CriticQ1Transitions, CriticQ2Transitions, PolicyTransitions);
|
| return;
|
| }
|
| Comment("");
|
| //--- Reject publication unless every requested iteration completed safely.
|
| if(completed_iterations != Iterations)
|
| {
|
| if(failed_iteration >= 0)
|
| PrintFormat("%s -> %d publication aborted: TrainTransition failed iteration=%d position=%d completed=%d/%d",
|
| __FUNCTION__, __LINE__, failed_iteration, failed_position,
|
| completed_iterations, Iterations);
|
| else
|
| PrintFormat("%s -> %d publication aborted: stop requested completed=%d/%d",
|
| __FUNCTION__, __LINE__, completed_iterations, Iterations);
|
| return;
|
| }
|
| if(!SavePolicies())
|
| PrintFormat("%s -> %d save failed", __FUNCTION__, __LINE__);
|
| else
|
| PrintFormat("ACSRM_STUDY_PASS iterations=%d critic=%I64u q1=%I64u q2=%I64u " +
|
| "q1_total=%I64u q2_total=%I64u policy=%I64u skip=%I64u teacher=%I64u " +
|
| "teacher_critic=%I64u random_critic=%I64u unresolved=%I64u " +
|
| "tags_tp=%I64u tags_sl=%I64u tags_horizon=%I64u " +
|
| "q1_err=%.8f q2_err=%.8f published=canonical",
|
| completed_iterations, CriticTransitions, CriticQ1Transitions, CriticQ2Transitions,
|
| CriticQ1Total, CriticQ2Total,
|
| PolicyTransitions, PolicySkippedInvalid, TeacherTransitions,
|
| TeacherCriticTransitions, RandomCriticTransitions, UnresolvedLabels,
|
| TagTP, TagSL, TagHorizon, CriticQ1Error, CriticQ2Error);
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Initializes the Stage 03 BasePolicy Expert. |
|
| //+------------------------------------------------------------------+
|
| int OnInit()
|
| {
|
| ResetLastError();
|
| //--- Pin the deal-evaluation horizon for manifest writer/validator symmetry.
|
| SkillManifestDealHorizon = MathMax(1, InpDealHorizon);
|
| SkillManifestTargetTau = double(InpTauT);
|
| SkillManifestTargetUpdatePeriod = InpUpdateTargets;
|
| if(InpSelector < 0 || InpSelector > 2 ||
|
| InpUpdateTargets < 0 || InpTauT < 0.0f || InpTauT > 1.0f)
|
| {
|
| PrintFormat("ACSRM input validation failed at line %d error=%d",
|
| __LINE__, GetLastError());
|
| return(INIT_FAILED);
|
| }
|
| //--- Confirm Stage 03 and that the frozen Forecast checkpoint is intact.
|
| if(InpOMPBStage != OMPB_STAGE_BASE_POLICY ||
|
| !SkillInitIndicators() || !SkillLoadForecastInference() ||
|
| !SkillConfigureProductionACSRMCheckpoint() ||
|
| !SkillValidateProductionACSRMCheckpoint() ||
|
| !SkillCaptureProductionACSRMFingerprints() || !LoadOrCreatePolicies() ||
|
| !SkillVerifyFrozenForecastExact() ||
|
| !SkillVerifyProductionACSRMFingerprints())
|
| {
|
| PrintFormat("ACSRM Actor-Critic initialization failed at line %d error=%d",
|
| __LINE__, GetLastError());
|
| return(INIT_FAILED);
|
| }
|
| if(!EventSetMillisecondTimer(1))
|
| {
|
| PrintFormat("ACSRM Actor-Critic timer initialization failed at line %d error=%d",
|
| __LINE__, GetLastError());
|
| return(INIT_FAILED);
|
| }
|
| //--- Finalize initialization only after the Stage 03 timer is armed.
|
| return(INIT_SUCCEEDED);
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Function OnDeinit. |
|
| //+------------------------------------------------------------------+
|
| void OnDeinit(const int reason)
|
| {
|
| EventKillTimer();
|
| //--- Only a completed explicit training loop persists all three artifacts
|
| //--- and writes the hash manifest last. Deinit must not create a mixed
|
| //--- generation.
|
| if(SkillForecast != NULL && !SkillVerifyFrozenForecastExact())
|
| PrintFormat("%s -> %d forecast mutation", __FUNCTION__, __LINE__);
|
| if(SkillProductionSignatureReady && !SkillVerifyProductionACSRMFingerprints())
|
| PrintFormat("%s -> %d ACSRM production signature mutation", __FUNCTION__, __LINE__);
|
| SkillForecast = NULL;
|
| }
|
| //+------------------------------------------------------------------+
|
| //| Runs the one-shot training lifecycle. |
|
| //+------------------------------------------------------------------+
|
| void OnTimer(void)
|
| {
|
| TrainACSRMActorCritic();
|
| ExpertRemove();
|
| }
|
| //+------------------------------------------------------------------+
|
