Seldonian Machine Learning Toolkit

Published:

Seldonian algorithms are machine learning algorithms that provide high-confidence guarantees that they will not produce undesirable behavior. The Seldonian framework provides guidance for the creation of useful Seldonian algorithms.

The Seldonian Toolkit provides basic Seldonian algorithms and interfaces, tools to facilitate the development of new Seldonian algorithms, and tools to evaluate the performance of Seldonian algorithms.

Contributions

  • Offline RL Algorithms: Implemented offline reinforcement learning algorithms like Weighted Importance Sampling (WIS), Per-Decision Importance Sampling (PDIS), and Doubly Robust for off-policy evaluation (OPE).
  • Sepsis Treatment Demo: Created a demo for the application of Seldonian RL algorithms for sepsis treatment in the ICU.

Website | Tutorials | GitHub