How tgosts/grnaproject works
Project Type and Purpose
• This is a standalone Java demo application (“Jogo demonstrativo com RNA”) combining a simple 2D game engine (package src.Jogo) with an in-house neural network library (package src.neuralNetwork).
• The game (player, enemies, shooting) leverages the neural network for AI decision-making or training scenarios, using external data files (src/data).
High-Level Architecture
The system naturally splits into four logical layers/modules:
a) Game Engine Module (src/Jogo)
b) Neural Network Module (src/neuralNetwork)
c) Data Module (src/data)
d) Utility/Configuration Module (src/util)
Main Components & Responsibilities
• Game Engine
• Classes: Ambiente, Jogador, Inimigo, Tiro, plus teste as an entry point.
• Responsibilities: game loop, rendering (if any), input handling, entity management.
• Interaction: At each tick, gathers state vectors, forwards them to the Neural Network Module for action decisions.
• Neural Network
• Classes: Rede (network), Camada (layer), Neuronio, Backpropagation, Funcao, FuncaoAtivacao.
• Responsibilities: model definition, forward propagation (inference), backpropagation (training), activation functions.
• Interaction:
– Training: reads data from Data Module via Properties config (paths to dataInTreino.txt), performs weight updates.
– Inference: receives real-time game state, outputs control signals for game entities.
• Data Module
• Files: dataIn.txt, dataInTreino.txt (both in src/data and mirrored in bin/data).
• Responsibilities: store raw feature vectors for training and testing.
• Interaction: consumed by Neural Network during initialization/training.
• Utility Module
• Class: Properties – centralizes configuration (e.g., learning rate, epoch count, file paths).
• Responsibilities: load .properties or custom config, pass parameters to Neural Network and Game Engine.
Relationships & Data Flows
• At startup, teste.java (or a main wrapper) loads configuration via util.Properties.
• neuralNetwork.Rede is instantiated with settings from Properties, optionally pre-trained or freshly initialized.
• If training mode is enabled, Backpropagation reads dataInTreino.txt → adjusts weights in Rede.
• Game loop (Ambiente) queries Rede with current sensor readings → obtains outputs for Jogador or Inimigo actions.
• Results propagate back to Ambiente, which updates game state.
Architectural Patterns & Design Principles
• Layered Architecture: clear separation between Presentation/Game Logic, Business Logic/AI, and Data Access.
• Single Responsibility: each package/class has a narrowly defined role (e.g., FuncaoAtivacao only implements activation functions).
• Dependency Injection via Properties: decouple hard-coded paths and parameters.