|

R194a
Empirical Evaluation of AND/OR Multivalued Decision Diagrams for Compilation and Inference

William Lam and Rina Dechter

Abstract
AND/OR Multi-valued Decision Diagrams (AOMDD) were shown provide a more compact representation of discrete-domain real- valued functions compared to other decision diagram variants. We show the performance of AOMDDs on compilation and inference tasks in graphical models. We introduce the elimination operator to AOMDDs, which in conjunction with the combination operator introduced in previous work, yields a full bucket elimination (BE) scheme using AOMDDs as an alternative function representation to tables. For compilation, we show that we can achieve a more compact AOMDD compared to previous work. For inference, we show that we are able to solve instances that do not fit in main memory when using tables.

[pdf]