Distributed Hypothesis Generation & Evaluation

This thesis addresses the challenges of modelling defeasible reasoning under uncertainty within intelligence analyses, focusing on hypothesis generation and evaluation -- an essential set of tasks where analysts must consider distributed, conflicting, uncertain, false, and potentially deceptive information. During analysis, analysts are required to generate the set of mutually exclusive and collectively exhaustive hypotheses to ensure that all the possible hypotheses are evaluated, and crucial hypotheses are not overlooked. With increasing data volumes, resulting from better collection methods and the advent of access to this data via the internet, the complexity of analysis, and the reasoning thereof, will increase also. This thesis is concerned with the development of artificial intelligence tools, techniques, and algorithms that provide analysts with reasoning-support during hypothesis generation and evaluation. The main output from this work is a number of proposals studying whether argumentation theory, computational argumentation, probability and information theory, and deep learning could be combined to offer assistance to analysts in these modelling tasks. Four techniques are presented and evaluated. The first approach extends a well-known argumentation scheme, deepening its level of specificity, and renders it capable of representing analysts' reasoning about hypotheses. The second proposal assesses the feasibility of using large language models for hypothesis generation. The third technique can refine an analyst's prior beliefs about the probability of hypotheses in light of the outputs from semantic evaluation of probabilistic argumentation frameworks. Finally, the fourth method employs abstract argumentation for sensitivity analysis and can identify arguments which, if found to be false or deceptive, would cause the most change to the conclusions drawn from an intelligence analysis. Throughout the thesis, both abstract and real-world examples are used to demonstrate the applicability of the novel techniques and algorithms presented. The proposals are proof of concepts and intended to be used as post-analysis quality checks, ensuring that analysts' reasoning is rational and includes all the information available to them.