Method for Real-Time Probabilistic Inference with Bayesian Network on GPGPU Devices

General Purpose Graphics Processing Unit (GPGPU) devices are used in most PCs for graphics, popular for high-performance computing, and relatively inexpensive. However, algorithms must be specifically designed for these devices.

The proposed invention consists of a process and a method for efficient probabilistic inference with Bayesian probabilistic networks on GPGPU devices. Bayesian probabilistic networks are widely used for modeling probability beliefs in computational biology and bioinformatics, healthcare, document classification, information retrieval, data fusion, decision support systems, security and law enforcement, betting/gaming and risk analysis.

 The invention consists of:

1. A novel “parallel irregular wavefront process” for importance sampling with Bayesian probabilistic networks, such that this process is tailored to the specific FPFPU device being used
2. A novel method to structure the Bayesian probabilistic network in GPGPU local memories to ensure optimal data access.

This invention increases the efficiency of probabilistic inference with Bayesian probabilistic networks on GPCPU devices. This is achieved by the specialized organization of data in the memory of these devices and by the optimized parallel process to produce results faster. The efficiency and performance increases commensurate with increasingly larger Bayesian probabilistic networks, i.e. the approach scales favorably with larger networks thereby making real-time probabilistic inference possible on large data sets and realistic applications.