Pacman Neural Network Classifier

A neural network built from scratch using NumPy to classify Pacman movement directions. No ML frameworks — forward pass, backpropagation, and weight updates are all implemented manually.

Architecture

Single hidden layer feedforward network:

Layer	Detail
Input	Game state feature vector (variable size)
Hidden	32 neurons, ReLU activation
Output	4 neurons (North, East, South, West), Softmax

Training

Parameter	Value
Learning rate	0.01
Epochs	500
L2 weight decay	0.01
Momentum	0.9

• L2 regularisation — penalises large weights during the backward pass to reduce overfitting
• Momentum — accumulates a velocity for each weight to smooth updates and help escape local minima
• Numerically stable Softmax — subtracts the row maximum before exponentiation to prevent overflow

Classes

NeuralNetwork — core implementation

• forward(X) — ReLU hidden layer → Softmax output, stores activations for backprop
• backward(X, y) — one-hot encodes targets, computes gradients with L2 reg, applies momentum updates
• fit(X, y) — runs the training loop
• predict(X) — returns the argmax class index

Classifier — Pacman integration wrapper

• fit(data, target) — converts training data to NumPy arrays and trains the network
• predict(features, legal) — runs inference and falls back to a random move if the predicted direction is illegal or the model hasn't been trained yet

Requirements

pip install numpy