Master's Thesis: Human-Centred Evaluation of Explainable Reinforcement Learning

Join our Team About this opportunity: Reinforcement Learning (RL) is increasingly used for sequential decision-making in areas such as robotics, recommender systems, autonomous control, and telecommunications. In telecom, RL is being explored for radio resource allocation, traffic steering, energy saving, and network self-optimisation. However, RL policies are typically opaque, making it difficult for operators, engineers, and researchers to understand why an agent selected a particular action. This is a serious obstacle to deployment in telecom, where trust, accountability, and the ability to diagnose misbehaviour are essential. Explainable Reinforcement Learning (XRL), a sub-field of Explainable AI (XAI), aims to make agent behaviour interpretable to humans. This thesis will design and run a user study comparing Feature Importance (FI) explanations with Temporal Policy Decomposition (TPD), which explains actions through predicted future outcomes. The study will investigate whether outcome-based explanations are more useful to humans than feature-attribution explanations in an RL context. The work corresponds to two students, 30 hp each, and can be organised into two subtracks. The students will collaborate on the user-study infrastructure and codebase. The location is Stockholm, Kista, and the preferred starting period is October 2026 to January 2027. What you will do:

• Review XAI and XRL literature and identify appropriate metrics and evaluation protocols for explanation quality.
• Design a user-study protocol based on four conditions:
• No explanation: participants see only the agent's actions.
• FI only: participants see feature-importance explanations.
• TPD only: participants see temporal-outcome explanations.
• TPD + FI: participants see both explanation types.
• Extend an existing web application to support the required XRL methods, the combined condition, and new measurements.
• Run a pilot study, refine the protocol, recruit participants, and conduct the main user study.
• Analyse the results statistically and evaluate the effectiveness of each explanation method individually and in combination.
• Write the thesis report and present the results to the research team.

The skills you bring:

• You are a Master's student in Computer Science, Human–Machine Interaction, Machine Learning, Data Science, or a related field.
• You have a foundation in machine learning, basic statistics, and data analysis.
• You have good programming skills in JavaScript and Python.
• You have good English proficiency and can communicate your findings clearly.