High-Performance MLP DLL for MQL5 (x64)

Version: 3.2 Author: Remind — Optimized Architecture Date: 2025

This repository contains a high-performance C++ Dynamic Link Library (DLL) implementing a Multi-Layer Perceptron (MLP) neural network. It is specifically designed for integration with MetaTrader 5 (MQL5). The library utilizes OpenMP for parallel processing, AVX/FMA instructions for mathematical precision, and the AdamW optimizer for efficient convergence.

Key Features

• AdamW Optimization: Implements the AdamW algorithm with decoupled weight decay for superior generalization compared to standard SGD or Adam.
• OpenMP Parallelism: Multi-threaded batch training capable of utilizing all available CPU cores.
• High Precision Math: Uses FMA (Fused Multiply-Add) and Neumaier summation to minimize floating-point accumulation errors during gradient descent.
• Stateless Architecture: Layers are designed to be memory efficient.
• Binary Persistence: Fast, compact binary format for saving and loading models.
• Built-in Visualization: Includes a lightweight GDI+ window to graph Mean Squared Error (MSE) in real-time.
• Loader Lock Safety: Version 3.2 addresses Windows Loader Lock deadlocks by implementing an explicit cleanup mechanism (NN_GlobalCleanup) instead of relying on DllMain detachment logic.

Prerequisites & Compilation

To build this project, you need a C++ compiler compatible with MSVC (Visual Studio 2019 or newer is recommended).

Project Settings

Ensure your project is configured with the following settings:

1. Architecture: x64 (Required for MQL5).
2. Language Standard: C++17 or later.
3. OpenMP Support: Enabled (/openmp). This is critical for performance.
4. Floating Point Model: Precise or Strict (recommended for Neumaier summation consistency).
5. Runtime Library: Multi-threaded DLL (/MD or /MDd).

API Reference

All functions are exported using extern "C" and the __cdecl calling convention.

Lifecycle Management

• int NN_Create() Allocates a new neural network instance. Returns a unique integer handle.
• void NN_Free(int handle) Frees a specific network instance.
• void NN_GlobalCleanup() CRITICAL: Must be called once before the MQL5 script/EA terminates. It safely shuts down the UI thread and releases all resources to prevent deadlocks.

Topology Configuration

• bool NN_AddDense(int handle, int input_size, int output_size, int activation) Adds a fully connected layer.
  • Activation Codes: 0=Sigmoid, 1=ReLU, 2=Tanh, 3=Linear, 4=Symmetric Sigmoid.
• void NN_SetSeed(unsigned int seed) Sets the RNG seed for weight initialization and shuffling.

Training & Inference

• bool NN_Forward(int handle, const double* input, int in_len, double* output, int out_len) Computes the forward pass for a single input vector.
• bool NN_TrainBatch(int handle, const double* inputs, int batch_size, int input_len, const double* targets, int target_len, double lr, double* out_mse) Performs a backward pass (training) on a batch of data using OpenMP.
  • inputs: Flat array containing batch_size * input_len elements.
  • out_mse: Optional pointer to receive the mean squared error of the batch.

Persistence

• bool NN_Save(int handle, const wchar_t* filename) Saves the model weights and topology to a binary file.
• bool NN_Load(int handle, const wchar_t* filename) Loads a model from a binary file.

Visualization (MSE Monitor)

• void NN_MSE_Show(int show) Show (1) or hide (0) the GDI+ graphing window.
• void NN_MSE_Push(double mse) Push a new error value to the graph.
• void NN_MSE_Clear() Clears the graph data.
• void NN_MSE_SetMaxPoints(int points) Sets the rolling window size for the graph.
• void NN_MSE_SetAutoScale(int enable, double min, double max) Enables autoscaling or sets fixed Y-axis bounds.

MQL5 Integration Example

Below is a snippet demonstrating how to properly import and manage the DLL within an Expert Advisor (EA).

// MLP_Bridge.mqh

#import "HighPerfMLP.dll"
   int  NN_Create();
   void NN_Free(int h);
   bool NN_AddDense(int h, int i, int o, int act);
   bool NN_TrainBatch(int h, const double &in[], int batch, int il, 
                      const double &tgt[], int tl, double lr, double &mse);
   bool NN_Forward(int h, const double &in[], int il, double &out[], int ol);
   
   // Critical cleanup function
   void NN_GlobalCleanup();
   
   // UI Functions
   void NN_MSE_Show(int s);
   void NN_MSE_Push(double m);
#import

// --- Expert Advisor Example ---

int net_handle = -1;

int OnInit() {
   // Initialize Network
   net_handle = NN_Create();
   if (net_handle <= 0) return INIT_FAILED;

   // Define Architecture: 10 Inputs -> 64 Hidden (ReLU) -> 1 Output (Sigmoid)
   NN_AddDense(net_handle, 10, 64, 1); // 1 = ReLU
   NN_AddDense(net_handle, 64, 1, 0);  // 0 = Sigmoid
   
   // Show Training Monitor
   NN_MSE_Show(1);
   
   return INIT_SUCCEEDED;
}

void OnDeinit(const int reason) {
   // CRITICAL: Call GlobalCleanup to stop threads safely
   // Do not rely on MQL5 to free the library automatically
   NN_GlobalCleanup();
}

void OnTick() {
   if (net_handle <= 0) return;

   // Example: Prepare dummy data (Replace with real market data)
   double inputs[10];
   double target[1];
   ArrayInitialize(inputs, 0.5);
   target[0] = 1.0;
   
   double mse = 0.0;
   double learning_rate = 0.001;
   
   // Train on a single sample (Batch size = 1)
   if (NN_TrainBatch(net_handle, inputs, 1, 10, target, 1, learning_rate, mse)) {
      NN_MSE_Push(mse);
   }
}