Multimodal explanations for explainable AI: an empirical study of text-video combinations on end-user understandability

Explainable Artificial Intelligence (XAI) has become increasingly important as deep reinforcement learning (RL) systems are deployed in safety-critical and human-facing applications. While recent methods provide post-hoc attributions for single decisions, they often fail to capture the sequential an...

Descripción completa

Detalles Bibliográficos
Autor: López Ortega, Anaïs
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/450577
Acceso en línea:https://hdl.handle.net/2117/450577
Access Level:acceso abierto
Palabra clave:Artificial intelligence
Reinforcement learning
Computer vision
Intel·ligència artificial
Aprenentatge per reforç
Visió per ordinador
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial
Descripción
Sumario:Explainable Artificial Intelligence (XAI) has become increasingly important as deep reinforcement learning (RL) systems are deployed in safety-critical and human-facing applications. While recent methods provide post-hoc attributions for single decisions, they often fail to capture the sequential and model-based nature of RL agents. In particular, model-based RL agents such as Dreamer rely on imagined rollouts to plan their actions, yet these internal simulations remain invisible to external observers. This creates a gap between what the agent predicts and what users can understand. This thesis investigates whether multimodal explanations, combining visual rollouts with natural language, can bridge this gap. A framework that integrates observed and imagined trajectories from DreamerV3 with instruction-tuned vision–language models (VLMs) is proposed. The pipeline extracts rollouts from MiniGrid environments, preprocesses them into single-frame, sequence, and comparative conditions, and generates explanations under a set of prompting strategies. After reviewing candidate models, Qwen-VL 2.5 was selected for its balance of accessibility, grounding quality, and feasibility. The resulting explanations were evaluated using a combination of automatic metrics and qualitative analysis focusing on factual consistency, temporal grounding, and divergence handling. The results show that VLMs can produce coherent descriptions of agent behaviour and partially capture temporal dependencies, though challenges remain with hallucinations and alignment between observed and imagined rollouts. Comparative explanations, in particular, highlighted divergences that revealed the agent’s planning process. This work contributes a reproducible methodology for generating and assessing multimodal explanations of model-based RL agents and suggests directions for future research, including user studies and integration with interactive explanation systems.