Interactive scheduling example
Retrain at the right time.
Choose how much data to use.
More training can improve a model, but it comes at a cost. Change the priorities and explore how the schedule responds.
Synthetic inputs · Exact offline planning · RetrainOpt
Loading precomputed schedules…
When + how much
Every bar is a retraining decision
The optimized plan can use 25%, 50%, 75% or 100% of available data. The three fixed policies use 100% whenever they retrain.
The setting
A changing stream
A toy regime value. Moving away from the model's training regime increases its loss.
The trade-off
Where the cost comes from
Inspect the numbers and assumptions
| Policy | Training | Raw loss | Weighted loss | Total | Retrains | Samples used |
|---|
These schedules are generated by the real RetrainingPlanner(...).plan("exact") using supplied synthetic costs. The optimizer sees all 24 steps in advance and minimizes the objective over the configured sample choices. This illustrates the decision problem; it is not a measured performance benchmark or a forecast.
Every policy starts with the same full-data model. “Every 6 steps” has a fixed interval, not an interval selected after seeing the results. No model is fitted for this demo; the loss and training-price formulas are documented in how it works.