Warrior_EA/AI_NETWORK.md

AI/ML Network Module Documentation

Overview

This document describes the architecture, design, and usage of the neural network and AI/ML infrastructure implemented in the Warrior_EA project. The core logic is implemented in AI/Network.mqh (MQL5) and accelerated with OpenCL kernels in AI/Network.cl.

---

1. Network.mqh (MQL5 Neural Network Framework)

Purpose

• Implements a modular, extensible neural network framework for use in trading signal generation, optimization, and adaptive money management.
• Supports multiple neuron types: standard, convolutional, pooling, and LSTM (recurrent) neurons.
• Provides both CPU and GPU (OpenCL) execution paths.
• Designed for integration with the EA's signal, filter, and scoring modules.

Key Components

• CConnection / CArrayCon: Represent weighted connections between neurons, with support for momentum and Adam optimizers.
• CNeuronBase / CNeuron / CNeuronPool / CNeuronConv / CNeuronLSTM: Hierarchy of neuron types, supporting feedforward, convolutional, pooling, and LSTM logic.
• CLayer / CArrayLayer: Layers of neurons, supporting flexible network topologies.
• CNet: The main neural network class, orchestrating layer construction, forward/backward propagation, and persistence.
• CNeuronBaseOCL: GPU-accelerated neuron implementation using OpenCL buffers and kernels.
• Enumerations: Activation functions (NONE, TANH, SIGMOID), optimizers (SGD, ADAM), buffer types.

Design Highlights

• OpenCL Integration: Uses #resource "Network.cl" for GPU acceleration. Automatically selects CPU/GPU path based on layer type.
• Extensible Neuron Types: Easily add new neuron types (e.g., attention, transformer) by extending the base classes.
• Persistence: Supports saving/loading network weights and structure to binary files for training and deployment.
• Dynamic Feature Configuration: Designed to allow dynamic selection of input features (price, volume, indicators) for ML pipelines.

Usage Patterns

• Construct network via CNet with a description array of CLayerDescription objects.
• Call feedForward() with input features, then backProp() with target outputs for training.
• Use Save()/Load() for persistence.

Known Issues & Recommendations

• Parameter Redefinition: Some parameters (e.g., eta, momentum) are defined both as macros and variables—refactor for clarity.
• Monolithic Structure: Consider splitting into smaller files by neuron type or functionality for maintainability.
• Unit Testing: No unit tests present—add tests for each neuron/layer type and for persistence logic.
• Documentation: This file now serves as the canonical reference for the AI/ML subsystem.

---

2. Network.cl (OpenCL Kernels)

Purpose

• Provides GPU-accelerated kernels for neural network operations: feedforward, output/hidden gradient calculation, and weight updates (Momentum/Adam).

Key Kernels

• FeedForward: Computes neuron activations for a layer.
• CaclOutputGradient: Calculates output layer gradients.
• CaclHiddenGradient: Calculates hidden layer gradients.
• UpdateWeightsMomentum: SGD with momentum.
• UpdateWeightsAdam: Adam optimizer.

Design Notes

• Uses double4 vectorization for performance.
• Handles activation functions (tanh, sigmoid) and edge cases for stability.
• Designed to be called from MQL5 via the OpenCL API wrappers in Network.mqh.

---

3. Integration Points

• Signals: Neural networks are used in AI-driven signal modules (e.g., PAI, CONV, LSTM signals).
• Database: Supports saving/loading trained weights for persistent learning.
• Configurable Features: Designed to allow dynamic selection of input features (price, volume, indicators).

---

4. Next Steps

• Refactor monolithic code into smaller, testable modules.
• Add unit and integration tests for all neural network components.
• Expand documentation for each neuron/layer type and kernel.
• Continue documenting all files in the workspace as per the project plan.

---

Workspace Documentation Progress

Documented Files

• AI/Network.mqh: Complete
• AI/Network.cl: Complete

Next Steps

• All files in AI/ are now documented.
• The following directories and files have been enumerated for documentation:
  • Database/
  • Enumerations/
  • Expert/
  • Money/
  • Signals/
  • Structures/
  • System/
  • Trailing/
  • Variables/
  • Warrior_EA.mq5
  • Warrior_EA.mqproj
  • README.md

Documentation will proceed systematically through each folder and file, updating this record as progress continues.

---

This section will be updated as each file is documented.