Context Mixing
Context mixing is the compression strategy used by the PAQ family: predictions from multiple independent context models are mixed with adaptive weights and fed into an arithmetic coder.
Purpose
It reaches extremely high compression ratios by ensembling predictions from many model types (orders, literal matches, neural networks, etc.).
Mechanism
Each model independently predicts the next symbol's probability; the mixer assigns weights adaptively based on past performance, and the weighted probability goes to the arithmetic coder. Mixers may be linear, logistic, or neural networks.
Use cases
- Extreme compression in the PAQ family
- Experimental compressors such as CMIX
- Compression competitions