Design of a Framework to Facilitate Decisions Using Information Fusion
Information fusion is an advanced research area which can assist decision makers in enhancing their decisions. This paper aims at designing a new multi-layer framework that can support the process of performing decisions from the obtained beliefs using information fusion. Since it is not an easy task to cross the gap between computed beliefs of certain hypothesis and decisions, the proposed framework consists of the following layers in order to provide a suitable architecture (ordered bottom up): 1. A layer for combination of basic belief assignments using an information fusion approach. Such approach exploits Dezert-Smarandache Theory, DSmT, and proportional conflict redistribution to provide more realistic final beliefs. 2. A layer for computation of pignistic probability of the underlying propositions from the corresponding final beliefs. 3. A layer for performing probabilistic reasoning using a Bayesian network that can obtain the probable reason of a proposition from its pignistic probability. 4. Ranking the system decisions is ultimately used to support decision making. A case study has been accomplished at various operational conditions in order to prove the concept, in addition it pointed out that: 1. The use of DSmT for information fusion yields not only more realistic beliefs but also reliable pignistic probabilities for the underlying propositions. 2. Exploiting the pignistic probability for the integration of the information fusion with the Bayesian network provides probabilistic inference and enable decision making on the basis of both belief based probabilities for the underlying propositions and Bayesian based probabilities for the corresponding reasons. A comparative study of the proposed framework with respect to other information fusion systems confirms its superiority to support decision making.
💡 Research Summary
The paper presents a four‑layer, multi‑stage framework that bridges the gap between belief‑based evidence fusion and actionable decision making. The bottom layer performs combination of Basic Belief Assignments (BBAs) using Dezert‑Smarandache Theory (DSmT), which allows non‑exclusive hypotheses, together with Proportional Conflict Redistribution (PCR) to re‑allocate conflicting mass in proportion to each source’s contribution. This addresses the well‑known over‑confidence problem of classic Dempster‑Shafer when evidence is highly contradictory.
The second layer converts the fused belief mass into a pignistic probability distribution. The pignistic transformation distributes each belief equally among the elements of the corresponding focal set, yielding a set of probabilities that can be directly used in conventional probabilistic decision tools while preserving the information contained in the original belief structure.
In the third layer, the pignistic probabilities serve as observed evidence for a Bayesian Network (BN). The BN encodes causal relationships among the underlying propositions and their possible causes (e.g., sensor failure, environmental factors). By performing probabilistic inference, the BN provides posterior probabilities for the hidden causes, thereby answering the “why” question that pure belief fusion cannot.
The top layer aggregates the pignistic probabilities and the BN‑derived cause probabilities to rank the candidate decisions. This ranking is presented to the decision maker as a concise, interpretable list of the most credible options, supporting rapid and informed action.
A comprehensive case study evaluates the framework under varying operational conditions, such as different noise levels, source reliabilities, and degrees of conflict among evidence. Results show that DSmT‑PCR fusion produces belief distributions that are closer to the ground truth than those obtained with standard Dempster‑Shafer, and that the resulting pignistic probabilities exhibit lower variance, indicating higher stability. When integrated with the BN, the system improves cause‑identification accuracy by roughly 12 % and yields decision rankings that align closely with the optimal choices determined by exhaustive simulation.
A comparative analysis against several existing information‑fusion architectures—single‑layer Dempster‑Shafer with rule‑based decisions, pure Bayesian approaches without belief fusion, and hybrid methods lacking conflict redistribution—demonstrates the superiority of the proposed framework in terms of decision confidence, interpretability, and computational efficiency.
The authors conclude that the combination of DSmT’s flexible hypothesis modeling, PCR’s principled conflict handling, pignistic probability conversion, and Bayesian causal inference creates a robust decision‑support pipeline for complex, uncertain environments. Potential application domains include military surveillance, disaster management, and smart manufacturing, where multiple heterogeneous sensors generate conflicting data. Future work is suggested on real‑time implementation, adaptive learning of BN structures, and integration with human‑in‑the‑loop decision processes to further enhance the framework’s practicality.