Using Artificial Neurons in Evidence Based Trust Computation

Çiçek Güven, Mike Holenderski, Tanir Özçelebi, Johan Lukkien

Abstract: This paper proposes an alternative approach for evidence based trust computation where the relationship between evidence and trust is learned using an artificial neuron, making it possible to automatically adapt trust computation to different use cases. Computational trust aims to quantify trust based on ever increasing evidence on observations. In the literature a trust value is seen as a posterior subjective probability, computed using Bayesian inference on evidence, a prior and a weight of the prior. This provides a fixed mapping between evidence and trust, which may not be suitable for every case study, e.g. when positive and negative evidences are not equally important. The proposed solution is also a first step towards our future work to replace complex and case-specific trust fusion operators proposed in the literature with a generic case-independent artificial neural network solution. Our experiments on example cases of coin toss prediction and occupancy detection show that for sufficiently large data sets, i.e. given sufficient evidence based on a history of observations, the proposed learning approach yields comparable results and in some cases beats the existing approach.