coarse coding

The general principle behind tile coding and RBFs. Instead of describing a state by one precise number, you describe it by which of many large, overlapping regions, called receptive fields, it falls in. Each region is one feature that fires for any state inside it. Because the regions are big and overlap, a single state lights up several features at once, and nearby states share most of them.

Generalization is set by the size and shape of the receptive fields, not by how many there are. Broad fields generalize widely (fast early learning, coarse final accuracy); narrow fields generalize little (slow but precise). Overlap is what makes the code coarse yet expressive: with enough overlapping fields, the combination of which features fire pins down the state finely even though each feature alone is blurry.

The lesson is counterintuitive — bigger features can mean faster learning, because each sample updates a broad swath of similar states, while final resolution stays high thanks to overlap. Tile coding is coarse coding with axis-aligned boxes; RBFs are coarse coding with smooth Gaussian fields.