Towards Tractable Dynamic Decision Making With Circuits

Gabriele Venturato, Vincent Derkinderen, Pedro Zuidberg Dos Martires, Luc De Raedt

Keywords: dynamic decision making, knowledge compilation, circuits, Markov decision processes, bayesian networks

Abstract:

A fundamental problem tackled by artificial intelligence is decision making under uncertainty in dynamic environments.
For example, a robot may need to autonomously reason on where to move (decision) at each time step (dynamic) while maximising the expected utility of the performed actions, and taking into account the inherent noisiness of the world (uncertainty).
Decision circuits have been shown to be a useful modelling tool in such settings, with the caveat that they do not treat time as a first-class citizen. We repair this omission by introducing dynamic decision circuits (DDCs). More specifically, we show how to obtain DDCs from dynamic decision-theoretic Bayesian networks via knowledge compilation and how to perform inference in DDCs using the algebraic model counting framework — a generalisation of weighted model counting.