A Neural Network Classifier Implementing Dempster-Shafer Evidence Theory

This document presents a neural network classifier grounded in Dempster-Shafer evidence theory. The classifier implements uncertainty quantification through belief functions and combines multiple evidence sources using Dempster's rule of combination. The architecture typically consists of an evidence extraction layer that converts input features into basic probability assignments, followed by an evidence fusion module that aggregates uncertainties through orthogonal summation. Key implementation aspects include defining frame-of-discernment structures, managing conflict between evidence sources using normalization factors, and implementing recursive combination algorithms for efficient computation. This hybrid approach leverages neural networks' pattern learning capabilities with Dempster-Shafer's mathematical framework for handling incomplete information, making it particularly effective for classification tasks with ambiguous or conflicting data. The system can be implemented using tensor operations for evidence combination and backpropagation for network optimization, achieving robust performance across various domains including medical diagnosis, fault detection, and multi-sensor data fusion scenarios.