When should locally weighted learning be used?

1. Always! We believe it is a good idea to approach all function approximation and data modeling first with locally weighted learning. It is a fast, efficient method of selecting features, validating models, estimating noise, and providing good insights into the relationships between variables.

2. What if a global parametric model can be obtained? Use the model locally and see Rule 1! Even if you have a global parametric model for your system, it often helps to fit it locally. Doing so can overcome inaccuracies in the global model caused by assumptions that only hold over a limited range of the state space.

3. What if the data set is extremely large? Buy more computer memory and processing and see Rule 1! The tree based algorithms used by Vizier make memory based learning efficient even in the face of large data sets. If the data set really is too big to keep around and process, it may be necessary to resort to other incremental learning algorithms that discard the data after training.

4. What if predictions must be very fast? Use caching methods and see Rule 1! When locally weighted learning produces accurate predictions, but faster speed is required, it is often useful to cache the predictions in a fast lookup scheme. The alternative is resorting to another function approximator that specializes in fast prediction times (neural networks can reach sub-millisecond prediction times).