Retraining Strategies: Scheduled, Triggered, and Online Learning

The problem, from first principles

03-model-degradation-retrain-trigger built the DECISION of whether a model needs retraining, but "when do we actually retrain?" has at least three genuinely different real answers, each with a different cost/responsiveness tradeoff: retrain on a fixed calendar schedule regardless of need, retrain only when degradation is actually detected, or never really "retrain" as a discrete event at all, instead nudging the model continuously with every new labeled example.

From theory to code

Implement scheduled_retrain_due, triggered_retrain_due (built directly on 03-model-degradation-retrain-trigger's has_model_degraded), and online_update_step, a genuine single-example gradient descent update.

Constraints

• scheduled_retrain_due(days_since_last_retrain, schedule_interval_days) returns True once the interval has elapsed, regardless of any performance signal.
• triggered_retrain_due(current_metric, baseline_metric, tolerance) returns exactly has_model_degraded's result.
• online_update_step(weight, bias, x, y, learning_rate) performs one step of gradient descent on the squared error (weight*x + bias - y)^2, returning the updated (weight, bias).
• The online update must always move (weight, bias) in a direction that DECREASES that squared error (for a small enough learning rate).

Hints

Open one at a time. Each gives away a little more than the last.