Computer Science

Permanent URI for this collection

https://hdl.handle.net/10530/366

Browse

Browsing Computer Science by Subject "Belief Change"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Belief change in probabilistic knowledge representations for open and dynamic computing environments
    (University of Zululand, 2020) Jembere, Edgar
    Theory refinement in Probabilistic Knowledge Representations is the task of up dating the Graphical Network structure in light of observations inconsistent with the current network structure. However, in the literature on Belief Change in Probabilis tic Knowledge Representations, theory refinement is only thought of as a change in the model parameters when data consistent with the Network structure is observed. Such Belief Change is not rich enough to capture the semantics of Belief Change in dynamic domains. In dynamic domains, the actual network structure at any given time is unknown and is unobservable. Only the data emitted from the domain is observable. Further to the foregoing, the Belief Change Model needs to cater to both changes necessititated by the correction of incorrect Beliefs (Belief Revision) and changes necessitated by changes in the domain (Belief Update). This thesis hy pothesised a Belief Change Meta-Model for Bayesian Network (BN) based Knowledge Representation in dynamic domains, and subsequently used the meta-model to de fine a Unified Belief Change Model for Bayesian Networks that caters for both Belief Revision and Belief Update of the Bayesian Network Structure. The Belief Change Model was conceptualised by first modelling the evolving Bayesian Network structure as a dynamical system whose impetus for change is driven by the occurrence of some events in the domain. The derived Unified Belief Change Model was formally validated by analogy using the Qualitative Belief Change Model for dynamic environments and the theory of Partially Observable Markov Decision Processes (POMDP). It was also proven that the proposed Belief Change model meets the postulates for revision of p-functions. xvii Apart from arguing the efficacy of the proposed Unified Belief Change Model from a theoretical standpoint, this thesis also provides empirical evidence for the same. A Belief Change operator, the Unified Belief Change Operator for Bayesian Networks (UBCOBaN), based on the proposed Belief Change Model was developed. The opera tor was then used to illustrate how the model achieves Belief Change using a synthetic example with one (1) iteration of Belief Change. Further to the fore-going the operator was implemented in java and was used for evaluating the efficacy of the model in both Propositional Bayesian Networks and in Multi-Entity Bayesian Networks (MEBN). MEBN is a variant of First Order Probabilistic Logic (FOPL) this research chose to use for evaluating the proposed model for Belief Change in First-Order Probabilis tic Knowledge Representations. The benchmark propositional Bayesian Networks used in the study were the ASIA, ALARM, HAILFINDER, HEPAR II, and the AN DES Bayesian Networks. The benchmark relational datasets considered for MEBN were the CORA, WebKP, UW std, and Financial std datasets. The results obtained showed that the proposed model adheres to the principle of minimal change (prin ciple information economy) better than the classical Search-and-Score algorithm in all the afore-mentioned propositional Bayesian Networks and all the datasets consid ered for MEBN. The model was also found to be at least as agile as the classical Search-and-Score algorithm in instances where data inconsistent with the assumed network structure was observed. This was observed for all the benchmark proposi tional Bayesian Networks used in ths study, and all the relational datasets consid ered for MEBN. The results obtained for an investigation on whether Belief Update improves rationality of the proposed Unified Belief Change Model on propositional Bayesian Networks showed that the Unified Belief Change Model with Belief Update has superior performance compared to the one without Belief Update. However, the superior performance was not statistically significant at 95% Confidence Interval.