Explainable AI for Decision Making

Abstract

This thesis examines the feasibility of using an AI system to support decision making processes in identifying potential root causes of quality issues in industrial and marine power systems. The AI system employs a Retrieval Augmented Generation (RAG) architecture, utilizing Large Language Models (LLMs). The research investigates whether pre-trained LLMs, combined with a constructed database in the RAG framework, are sufficient to provide support in a highly specific domain context. It also explores the factors that influence user acceptance and trust in the AI system. The evaluation includes both quantitative metrics and qualitative user tests with domain experts. The project was conducted in collaboration with Volvo Penta, a power solution provider, and all data collection and user testing were performed at the company. The findings suggest that the system can effectively retrieve and summarize historical data to aid in identifying the root causes of quality issues. Additionally, the study reveals that user satisfaction and trust of AI-driven insights are primarily influenced by the system’s ability to explain its reasoning process for reaching conclusions.