Even better Open-source AIT Software: AIQI, larger model class, Python API + more

[Noah]

Hello everyone, I previously shared Infotheory, an Algorithmic Information Theory library and CLI. 

I've just released a major update, which extends it's usefulness, makes it portable and greatly improves performance and memory efficiency broadly.

I provide an explanation of the new features and capabilities, reproducible benchmark results, which indicate Infotheory's MC-AIXI-CTW is faster, and uses less memory across all tested configurations compared to PyAIXI and the Reference C++ Implementation of MC-AIXI-CTW. 

Infotheory aims to unify sequential probabilistic models interfaces, which are plug-compatible with metric functions, search, agents, compression, generation and more, and states can be managed across those different use-cases.

For those which learn best by doing something, you may try the new demos, which now include in-browser MC-AIXI and AIQI: https://infotheory.tech  and click AIXI Runner -- no data leaves your web browser, and everything runs locally with WebAssembly. It's fast but single threaded, which exemplifies how much faster AIQI is. 

Infotheory now has one main model class: Probablistic Predictive models ("Rate Backends") -- which can be used to construct compressors(Add AC or rANS). This is a major change from v1.0.0

I've added new models (Mamba v1 LLM, Match, PPMd), improved performance and memory usage of existing models including CTW, added generative capability, added an AIQI approximation to compliment MC-AIXI, made all LLM models capable of Online training+weight export in addition to inference on weights.

In summary:

1. It now has a Python API, allowing complete usage of all features with Python alone without knowing Rust (which the underlying library is written in) -- including AIXI, and yes, environments can be defined in Python. Just `pip install infotheory-rs` --  Documentation is available on the site.

2. It is now completely portable across operating systems and architectures -- even web browsers via WebAssembly, which is very performant. 

3.  Extended Model class significantly -- models work universally within the library, plug and play, and can be arbitrarily and recursively configured in mixtures, of there is now also Neural Mixture (ZPAQ inspired) 

4. Implemented AIQI-CTW, and extended it with any library model as the return predictor! Infotheory's AIQI-CTW yields a general 5-10x improvement in throughput compared to it's own MC-AIXI-CTW implementation; See my later benchmark section.

5. Any Infotheory Model can be used for compression and decompression(serving as a good evaluator), losslessly. Since models can be Mixtures(Bayesian, Neural) of other models or mixtures, which themselves are configurable -- It's a compression framework. You configure it down to what algorithms are used and how.

In addition, I have improved the documentation, and updated my site from a server architecture, to allowing you to demo it's features inside your web browser locally via WebAssembly. The demo has a few basic games, and you can watch the Agent play, it's very fun to tweak and play around with in my opinion.

Benchmarking:

Infotheory's MC-AIXI-CTW implementation is 70-140x faster than PyAIXI's in equivalent scenarios.

It is likewise 1.4-3x faster than this C++ MC-AIXI implementation in the same equivalent scenarios, never slower. It consistently uses less memory compared to both implementations, strictly. I observed strictly superior speed and memory efficiency. 

Across 18 AC-CTW scenario-trials: reward for Infotheory's MC-AIXI was below both in 6, above both in 5, between them in 7.

The exact steps to reproduce these results are documented on my site (the entire thing is a deterministic computation)

Hardware is Intel(R) Core(TM) i7-8850H and all runs were single-threaded, and ran on Linux.

Infotheory supports efficient, full planner parallelism for both AIQI and MC-AIXI, which further improves performance.

And Infotheory's AIQI-CTW is 5-10X faster, and more memory efficient than even its own MC-AIXI implementation, which confirms and reproduces the findings made by Yegon Kim and Juho Lee that AIQI-CTW can be implemented faster than MC-AIXI-CTW, and I confirm it can be parallelized to practically the same degree. 

In the attached graphs, infotheory-rust refers to the Infotheory CLI, whereas Infotheory-python is Infotheory used via it's Python API, which compute identical rewards. 

I will be releasing more examples, documentation, and benchmarks on my site including Jupyter notebooks in the coming days... 

I would very much welcome feedback and suggestions from the community. 

Best,

Noah