Indirect Causes in Dynamic Bayesian Networks Revisited
Authors: Alexander Motzek, Ralf Möller
JAIR 2017 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide substantiating empirical evidence for the latter results by showing that solving multiple onand offline filteringand smoothing-problems remains tractable even over large periods of time. All experiments are reproducible, for which we supply an experimental framework implementing inference in dense ADBNs in C available at: http://adbn.motzek.org |
| Researcher Affiliation | Academia | Alexander Motzek EMAIL Ralf Möller EMAIL Universität zu Lübeck Institut für Informationssysteme 23562 Lübeck, Germany |
| Pseudocode | Yes | Algorithm 1 Evolving a Sample ... Algorithm 2 Sequential Importance (Re)Sampling (SIS/SIR) |
| Open Source Code | Yes | All experiments are reproducible, for which we supply an experimental framework implementing inference in dense ADBNs in C available at: http://adbn.motzek.org |
| Open Datasets | No | For every timestep t > 0 we generate random observations bt conforming with Theorem 1. We refrain from observing state variables Xt, i.e., z t = , in order to achieve worst-case time complexity. ... In every experiment, we consecutively solve the online filtering problem... |
| Dataset Splits | No | The paper describes generating random data and observations for simulations and experiments, but does not involve standard datasets with predefined training, validation, or test splits. The evaluation focuses on computational complexity and accuracy of inference methods over simulated data. |
| Hardware Specification | No | The paper discusses computation time in milliseconds and mentions limits like 250GB of memory for exact inference, but it does not specify any particular CPU, GPU, or other hardware models used for running the experiments. |
| Software Dependencies | No | All experiments are reproducible, for which we supply an experimental framework implementing inference in dense ADBNs in C available at: http://adbn.motzek.org. While the language 'C' is mentioned, no specific compiler version or library versions are provided. |
| Experiment Setup | Yes | To do so, we perform multiple experiments, where in every experiment, state variables in Xt, t > 0 are assigned a randomly generated individual CPD following Definition 2. Further, in every experiment random priors are assigned to random variables X0 and At. Every randomly generated probability is taken from the range [0.1, 0.9] to avoid impossible observations. ... Experiments were repeated 139 times for n = 4, i.e., an ADBN consisting of 16 random variables Xt, At per timeslice for a timerange of 40. ... For n = 4 one state variable Xt i is observed per timestep and for n = 5 two are observed. ... For finite amounts of samples and finite numerical precision, SIS and SIR deliver approximate solutions for filtering problems. However, for finite amounts of samples, SIS techniques suffer from a degeneracy problem (cf., Murphy, 2012, pp. 823 831; Doucet & Johansen, 2009) in DBNs, as well as in ADBNs: Gradually, a significant amount of samples degrades and carries a very low weight... |