Capacity and regularisation · Data Playground

Understand the idea

Width increases representational capacity. Alpha controls weight regularisation. Use controlled training/validation evidence to reason about both; production’s small width grid does not explore every possible network.

Interpret controlled width and alpha experiments.Scroll the diagram horizontally if needed.

What the syntax does

hidden_layer_sizes=(16, 8): Requests two hidden layers, with 16 units and then 8 units.
hidden_layer_sizes: (24,), (16, 8)
alpha=0.01: Sets the L2 weight penalty; its useful value needs validation evidence.
alpha: 0.0001, 0.01
model.set_params: Changes estimator settings.
Refit before evaluating the changed model.

Outputs you create

layers: Requested result described in the task and deliverables.
alpha: Requested result described in the task and deliverables.

LINE24 · first 8 rows
`distance`	`duration`
1	5.30501
1.47826	11.2361
1.95652	11.8488
2.43478	14.809
2.91304	10.7589
3.3913	19.4972
3.86957	17.89
4.34783	15.531

Your task

Create a network with one 24-unit hidden layer and alpha=0.0001.
Change its hidden layers to (16, 8) and alpha to 0.01.
Store the resulting settings in layers and alpha.

layersalpha

Hints

Think: Use set_params(), tuples and a float to describe capacity and regularisation before fitting.

Tools: Use hidden_layer_sizes=(16, 8), alpha=0.01, model.set_params. Read the visible syntax meanings before editing.

Approach: Create a network with one 24-unit hidden layer and alpha=0.0001. Change its hidden layers to (16, 8) and alpha to 0.01. Store the resulting settings in layers and alpha. Keep the supplied row order and inspect the named output after running.

Explained solution

from sklearn.neural_network import MLPClassifier
model = MLPClassifier(hidden_layer_sizes=(24,), alpha=0.0001, random_state=42)
model.set_params(hidden_layer_sizes=(16, 8), alpha=0.01)
layers = model.hidden_layer_sizes
alpha = model.alpha

The new tuple changes the network capacity and alpha changes the penalty on large weights. These are candidate settings, not automatic improvements. Compare candidates on the same training folds and consider convergence and cost.