Multi-layer Perceptron

Neural network with fully connected layers for regression.

When to use:

• Complex non-linear patterns
• Large datasets
• Modern alternative to traditional ML
• Can sacrifice interpretability

Strengths: Handles very complex patterns, flexible architecture, proven in production Weaknesses: Needs more data, longer training, requires tuning, black box

Model Parameters

Hidden Layer Sizes (default: (100,)) Neurons in each hidden layer. Example: (100, 50) = 2 layers with 100 and 50 neurons.

Activation

• relu: Rectified Linear Unit (default, most common)
• tanh: Hyperbolic tan (smooth)
• logistic: Sigmoid function
• identity: Linear (for linear regression)

Solver

• adam: Adaptive moment estimation (default, best for large data)
• sgd: Stochastic gradient descent
• lbfgs: Quasi-Newton (good for small data)

Alpha (default: 0.0001) L2 regularization parameter.

Learning Rate

• constant: Fixed learning rate
• invscaling: Gradually decreasing
• adaptive: Adapts based on performance

Max Iterations (default: 200) Maximum training epochs.

Early Stopping (default: false) Stop training when validation score stops improving.

Random State (default: 42) Seed for reproducibility.