Decision Trees Explain Finite-Memory Policies for POMDPs

A new method from arXiv:2411.13365v2 represents finite-memory policies for Partially Observable Markov Decision Processes (POMDPs) using decision trees and Mealy machines to improve explainability. POMDPs are a key framework for decision-making under uncertainty, but optimal policies often require infinite memory, making them hard to implement and undecidable. Finite-memory policies are more practical, yet their algorithms and resulting policies remain complex. The proposed approach combines Mealy machines, which switch between stationary parts, with decision trees that describe simple, stationary components. This yields policies that are both interpretable and smaller, enhancing explainability. The translation applies to policies in the finite-state-controller (FSC) form from standard literature.