gavindiaz/TR_Agent
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix sym shadowing warnings + coherent softmax gating

Browse Source 
- Rename Init() param sym -> symName in IAgent base and overrides
  (MAAgent, MomentumAgent, RegimeADX) to stop hiding the global sym.
- Softmax gating: derive temperature from dispersion of the competence
  weights ws[] (the same quantity the softmax weights), not from the
  variance of signals zs[]; skip disabled agents in the normalization.
- Use |rho| (with weightMin floor) as competence and sign-flip
  anti-correlated agents instead of zeroing them, so rho<0 information
  is used (inverted) rather than discarded.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
This commit is contained in:
Pietro Giacobazzi
2026-06-14 18:20:19 +00:00
parent 3256ebb5c6
commit a8b4a7d3ff
5 changed files with 37 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
MQL5/Experts/MultiAgentTest/Core/Orchestrator.mqh
+29 -16
View File
@@ -375,14 +375,18 @@ public:
double Analyze(const MarketData &data) {
if(agentCount == 0) { combinedZ = 0; return 0; }
double zs[MAX_AGENTS], ws[MAX_AGENTS];
double zs[MAX_AGENTS], ws[MAX_AGENTS], corrSign[MAX_AGENTS];
// Fase 1: Analisi individuale
for(int i=0; i<agentCount; i++) {
if(!agents[i].enabled) { zs[i] = 0; ws[i] = 0; continue; }
if(!agents[i].enabled) { zs[i] = 0; ws[i] = 0; corrSign[i] = 1.0; continue; }
zs[i] = agents[i].Analyze(data);
double rho = GetCorrelation(i);
ws[i] = (rho > 0) ? MathMax(weightMin, rho) : 0;
// Competenza = |correlazione| segnale/ritorni (con floor weightMin, così
// ogni agente contribuisce una baseline). Il segno di rho non viene scartato:
// un agente anti-correlato è informativo → il suo segnale viene invertito.
ws[i] = MathMax(weightMin, MathAbs(rho));
corrSign[i] = (rho < 0) ? -1.0 : 1.0;
}
// Fase 2: Interazione tra agenti (es. RegimeDetector aggiorna shared context)
@@ -419,32 +423,41 @@ public:
combinedZ = m_neuralNet.GetCombinedZ(nnInput);
combinedZ = MathTanh(combinedZ);
} else {
// Classic softmax gating (existing logic)
double zMean = 0, zVar = 0;
int zCount = 0;
// Softmax gating coerente: il softmax pesa le competenze ws[] e la
// temperatura deriva dalla dispersione delle STESSE competenze (non
// dalla varianza dei segnali zs[]), così i due termini sono consistenti.
// Competenze simili → temp bassa; competenze molto diverse → temp alta
// (mixing più uniforme, evita di sovra-fidarsi di un singolo agente).
double wMean = 0;
int wCount = 0;
for(int i=0; i<agentCount; i++) {
if(!agents[i].enabled) continue;
zMean += zs[i];
zCount++;
wMean += ws[i];
wCount++;
}
if(zCount > 0) zMean /= zCount;
if(wCount > 0) wMean /= wCount;
double wVar = 0;
for(int i=0; i<agentCount; i++) {
if(!agents[i].enabled) continue;
zVar += (zs[i] - zMean) * (zs[i] - zMean);
wVar += (ws[i] - wMean) * (ws[i] - wMean);
}
if(zCount > 0) zVar /= zCount;
double temp = (zCount > 0) ? (MathSqrt(zVar) + DATA_EPS(zVar)) / MathSqrt(zCount) : 1.0 / MathSqrt(MathMax(1, agentCount));
if(wCount > 0) wVar /= wCount;
double temp = (wCount > 0) ? (MathSqrt(wVar) + DATA_EPS(wVar)) / MathSqrt(wCount) : 1.0 / MathSqrt(MathMax(1, agentCount));
double safeTemp = MathMax(temp, DATA_EPS(temp) * 10.0); // evita overflow di MathExp
double sumExp = 0;
for(int i=0; i<agentCount; i++)
for(int i=0; i<agentCount; i++) {
if(!agents[i].enabled) continue;
sumExp += MathExp(ws[i] / safeTemp);
}
combinedZ = 0;
double epsExp = DATA_EPS((double)agentCount);
double epsExp = DATA_EPS((double)wCount);
for(int i=0; i<agentCount; i++) {
double softmaxW = (sumExp > epsExp) ? MathExp(ws[i] / safeTemp) / sumExp : 1.0 / agentCount;
combinedZ += softmaxW * zs[i];
if(!agents[i].enabled) continue;
double softmaxW = (sumExp > epsExp) ? MathExp(ws[i] / safeTemp) / sumExp : 1.0 / MathMax(1, wCount);
// Segnale sign-corretto: agenti anti-correlati contribuiscono invertiti.
combinedZ += softmaxW * corrSign[i] * zs[i];
}
combinedZ = MathTanh(combinedZ);
}