dmorrill10 / seq_predict

Module for sequence prediction and compression using Bayesian prediction models.

Sequence Prediction

Copyright (c) 2016 Michael Bowling [e-mail]

See LICENSE for information on using or distributing the code.

Module for sequence prediction and compression using Bayesian prediction models.

Implements the following Bayesian models:

• KT
• SAD
• CTW
• CTS
• PTW
• An early version of Forget-Me-Not
• Model combinations: Averaging, Factoring

Most of the implementations are in pure python. CTW and CTS with integer k-ary alphabets (i.e., 0, 1, ..., k) using KT estimators has a fast c-based implementation.

In addition, it includes an arithmetic binary coder.

Using it for Compression

Executing the module on the command line invokes a compression tool that allows you to select one of the implemented models to use with an arithmetic encoder for compression.

python -m seq_predict compress FILE.TXT FILE.TXT.compressed -m CTW

Using it for Prediction

The model classes can be imported as below.

from compress.models import PTW
from compress.fast import CTS_KT

The models adhere to the following API:

• predict(symbol, history)
• log_predict(symbol, history)
• log_predict_seq(symbols, history)
• update(symbol, history)
• update_seq(symbols, history)
• copy()

The history is optional for models that don't look at history (e.g., KT estimators). The *_seq functions only make sense if appending symbols are how the history is updated.