Credal Marginal MAP

Authors: Radu Marinescu, Debarun Bhattacharjya, Junkyu Lee, Fabio Cozman, Alexander Gray

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental An extensive empirical evaluation demonstrates the effectiveness of our new methods on random as well as real-world benchmark problems. We evaluate empirically the new CMMAP inference algorithms on random credal networks with different graph topologies as well as a collection of credal networks derived from real-world applications.
Researcher Affiliation Collaboration Radu Marinescu IBM Research, Ireland radu.marinescu@ie.ibm.com Debarun Bhattacharjya IBM Research, USA debarunb@us.ibm.com Junkyu Lee IBM Research, USA junkyu.lee@ibm.com Alexander Gray IBM Research, USA alexander.gray@ibm.com Fabio Cozman Universidade de São Paulo, Brazil fgcozman@usp.br
Pseudocode Yes Algorithm 1 Variable Elimination for Credal Marginal MAP; Algorithm 2 Mini-Buckets for Credal Marginal MAP; Algorithm 3 Local Search for Credal Marginal MAP
Open Source Code Yes The supplementary material includes additional details, experimental results, code and benchmarks.
Open Datasets Yes We evaluate the proposed algorithms for CMMAP on random credal networks and credal networks derived from real-world applications. ... credal networks derived from 22 real-world Bayesian networks1 ... 1Available at https://www.bnlearn.com/bnrepository/
Dataset Splits No The paper does not provide specific details on training, validation, and test dataset splits (percentages, counts, or specific methods like k-fold cross-validation).
Hardware Specification Yes All competing algorithms were implemented in C++ and the experiments were run on a 32-core machine with 128GB of RAM running Ubuntu Linux 20.04.
Software Dependencies No The paper mentions 'C++' and 'Ubuntu Linux 20.04' but does not specify version numbers for any libraries, solvers, or other key software dependencies.
Experiment Setup Yes The local search algorithms used N = 10 iterations and M = 10, 000 maximum flips per iteration, and they all used the approximate L2U algorithm with 10 iterations [29] to evaluate the MAP assignments during search. Furthermore, for SHC we set the flip probability pflip to 0.2, TS used a taboo list of size 100, while for SA we set the initial temperature and cooling schedule to Tinit = 100 and σ = 0.9, respectively. For CMBE(i) we set the i-bound i to 2 and used the same L2U algorithm to evaluate the solution found. All competing algorithms were allocated a 1 hour time limit and 8GB of memory per problem instance.