Neural Network (Java)

Lightweight multi-layer perceptron implementation written in Java, focused on teaching-grade experiments and small reinforcement-learning prototypes.

Requirements

• JDK 8 or newer (project targets bytecode level 8 by default)
• Bash-compatible shell for the helper scripts

Project Layout

• src/mg/rivolink/ai – core network, layers, and neuron primitives
• src/mg/rivolink/io – model persistence utilities (binary, JSON, XML wrappers)
• src/mg/rivolink/test – runnable smoke tests that cover XOR/AND/OR training scenarios
• scripts/ – build tooling (build.sh)
• dist/ – packaged jars after a build
• models/ – persisted models produced by IO tests or manual experiments

Build & Artifacts

make build

Invokes scripts/build.sh --clean, compiles the library, and emits dist/neural-network.jar plus a matching sources jar.
Use ./scripts/build.sh --target=11 if you need a different Java release.

Running Tests & Demos

Compile then execute any of the main-based tests:

javac -cp src -d bin $(find src -name '*.java')
java -cp bin mg.rivolink.test.Network1HiddenTest
java -cp bin mg.rivolink.test.Network2HiddenTest
java -cp bin mg.rivolink.test.NetworkIOTest

Each test prints convergence diagnostics or IO verification results to stdout.

Saving & Loading Models

Use NetworkIO.save(network, path) for binary models, or pass an explicit format:

NetworkIO.save(network, "models/xor.json", NetworkIO.Format.JSON);
Network restored = NetworkIO.load("models/xor.json");

Binary (.bin), JSON (.json), and XML (.xml) files are interchangeable and round-trip-compatible via the built-in serialization wrappers.