Ace is a package that compiles a Bayesian network into an Arithmetic Circuit (AC) and then uses the AC to answer multiple queries with respect to the network. Ace's approach to probabilistic inference has two chief advantages. First, it pushes much of the work involved in performing inference to an offline phase, which can then be amortized across many online queries. Second, it makes highly effective use of certain types of local structure (especially determinism) in the network in addition to global structure, to make inference more efficient.
Ace version 3.0 was released in August 2015 and contains many improvements.
See the included documentation for details.
Ace as an encoder
Compilation proceeds by encoding the network into CNF, compiling the CNF into d-DNNF (using the c2d knowledge compiler), and extracting the AC from the compiled d-DNNF. Ace can also be used to encode networks into CNF without compiling, which makes it a valuable source of interesting problems for weighted model counters and knowledge compilers.
Download Ace 3.0 and/or benchmarks.
Obtain additional information about Ace.
View some additional results (using Ace v1.0).
Send questions and comments to ace at cs.ucla.edu.