Classical neural networks and game AI implemented from first principles in C#. No ML libraries: every neuron, weight update, and search tree here is written by hand. The goal was to understand the math by building it, the same philosophy as my custom CPU and toolchain: work one layer below the abstraction you normally get for free.
| Project | What it is | Key files |
|---|---|---|
| Perceptron | Single-layer perceptron with mutation-based learning, plus a line-of-best-fit regression demo. Activation, error, normalization, and random-init functions are split into swappable modules. | NeuralNetworks/Perceptron/Perceptron.cs, LineOfBestFitPerceptron.cs, ActivationFunctions.cs, ErrorFunctions.cs |
| Feed-Forward Neural Network | A multi-layer network built object by object: neurons hold dendrites, dendrites hold weights, layers hold neurons. Nothing is a matrix until you understand why it should be. | NeuralNetworks/Feed Forward neural Network/Neuron.cs, Dendrite.cs, Layer.cs, FFNN.cs |
| Backpropagation | Console demo that trains the feed-forward network with backpropagation (gradient descent via the chain rule). | Backpropogation/Program.cs |
| Hill Climber | A black-box optimizer: mutate, keep improvements, discard regressions. Useful as a baseline to appreciate what gradients buy you. | NeuralNetworks/Hill Climber/HillClimber.cs |
| MiniMax | Game-tree search with alpha-beta pruning. | MiniMax/Program.cs |
| Genetic Flappy Bird | A population of networks learns Flappy Bird through neuroevolution, rendered live in MonoGame. Watch generations go from faceplanting into the first pipe to playing indefinitely. | Flappy Game/Bird.cs, Game1.cs, PipeManager.cs |
The feed-forward implementation deliberately avoids matrix shortcuts. A Neuron owns its incoming Dendrite connections, each Dendrite owns a weight, and a Layer is a collection of neurons. A forward pass walks this object graph. Training happens two ways: the hill climber mutates weights blindly and keeps what works, while the backpropagation project computes exact gradients and follows them. Implementing both made the difference between them concrete: one is search, the other is calculus.
Requires the .NET SDK. Each project is a standalone console app inside one solution (NeuralNetworks.sln), so you can also open everything in Visual Studio or Rider.
git clone https://github.com/Aidan-Gildea/NeuralNetworks.git
cd NeuralNetworks
dotnet run --project Backpropogation # train the MLP with backprop
dotnet run --project MiniMax # game-tree search demo
dotnet run --project "Flappy Game" # genetic Flappy Bird (MonoGame window)The Flappy Bird project restores MonoGame automatically through NuGet on first build.
- full-stack-8bit-cpu: an 8-bit CPU built gate by gate in Logisim, plus the custom ISA, assembler, and emulator that program it.
- robot-arm-3dof: a 3-DOF robot arm with inverse kinematics derived from first principles.
MIT