A Fuzzy-Dynamic Bayesian Network Approach for Inference Filtering

Bayesian Networks (BN) are used for representing and inferring over variables with aleatory uncertainty. Dynamic Bayesian Networks (DBN) extend this concept by introducing temporal dependencies that catch dynamic behaviors from the domain variables. Effective and efficient modeling through BN demands data discretization on categories. However, these categories may have vagueness uncertainty, once are used labels not defined by exact numerical thresholds. Fuzzy Theory provides a framework for modeling vagueness uncertainty. Although hybrid theories to integrate Fuzzy Theory and BN inference process have been proposed, there are still limitations on using fuzzy evidence on DBN. The related works restrict the evidence modeling to the overlapping of only two fuzzy membership functions. Thereby, this work proposes a method for Dynamic Fuzzy-Bayesian inference over non-dichotomic variables. To evaluate the proposal, the model is applied as a classifier on the Detection Occupancy Dataset and compared with other approaches. In the experiments, the model obtained Accuracy 97% and Recall 92%.