Combining A Fortiori Reasoning and a Similarity Measure in Case-Based Reasoning

Abstract

Over the last years, case-based reasoning has shown to be a promising AI-method in the legal domain, in which a fortiori reasoning is utilized to find similar cases from the past. However, this approach limits the decision making process by not being able to make a prediction for every new case, resulting in a significant number of cases remaining undecided. This thesis discusses the development and evaluation of a newly created case-based reasoning (C-BR) model designed to address this issue, by combining two previously designed models: the aforementioned legal C-BR model and a traditional similarity measure-based C-BR model. Similarity measures are commonly employed in C-BR models to retrieve previous cases in order to solve new problems. The combined approach proposed in this case study includes a fortiori reasoning as well, which involves a formal model of legal reasoning to retrieve similar cases from the past. This combined approach aims to improve the accuracy and reliability of C-BR models. The combined model was tested on a decision making problem of the CBR (‘Centraal Bureau Rijvaardigheidsbewijzen’: Dutch Central Office of Driving Certification), in which a decision of the fitness to drive of individuals was made based on their health deviations. The performance of the combined C-BR model was compared to models solely utilizing one of the two approaches. While results indicate that the combination of both techniques is a promising approach in C-BR, this model does not outperform the more traditional approach yet, making less humanlike decisions than traditional C-BR with a similarity measure. However, this study emphasizes the necessity of further research into the applicability of the integrated model in other domains. Future studies could investigate the potential of the combined C-BR model in datasets that are less complex. This could increase the performance of case-based reasoning models in AI, making these models applicable in even more domains to assist human decision making.