Implementation of an Incremental MDL-Based Two Part Compression Algorithm for Model Inference

Abstract

We describe the implementation and performance of a compression-based model inference engine, MDLcompress. The MDL-based compression produces a two part code of the training data, with the model portion of the code being used to compress and classify test data. We present pseudo-code of the algorithms for model generation and explore the conflicting requirements between minimizing grammar size and minimizing descriptive cost. We show results of a MDL model-based classification system for network traffic anomaly detection.