Reward Design Without the Expert: Testing LLM Assistance for Multi-Agent RL in Electricity Markets

Research topic/area: Reinforcement Learning · Large Language Models · Empirical IS Research
Type of thesis: Bachelor / Master
Start time: -
Application deadline: 31.08.2026
Duration of the thesis: ab sofort

Description

Reward Design Without the Expert: Testing LLM Assistance for Multi-Agent RL in Electricity Markets

Background

Reinforcement learning is increasingly used to simulate strategic bidding in electricity markets, to test market designs before deployment or to study actors under new regulation. The bottleneck is reward design: it requires both deep market knowledge and RL experience, a rare combination that limits who can use Multi-Agent Reinforcement Learning (MARL) for market analysis today.

LLMs can generate reward functions from natural-language descriptions (Kwon 2023; Eureka; Text2Reward; REvolve 2024). But existing evaluations cover only robotics and game environments with technical metrics. Whether LLM support can substitute for or complement domain expertise, and whether it works in electricity-market MARL, has not been empirically tested.

What You Will Do

You will work directly with ASSUME, an open-source MARL framework for agent-based electricity-market simulation, and build a reward specification system that translates natural-language inputs into executable reward functions for heterogeneous market agents. The prototype is then evaluated in a user study comparing experts and non-experts in reward design tasks.

Expected Contributions

Methodological: First systematic test of whether LLM support reduces the expertise requirement in MARL reward design.

Technical: A working prototype for stakeholder-specifiable reward design on ASSUME.

Empirical: Quantitative and qualitative evidence on democratization effects in human–LLM collaboration for complex design tasks.

Disclaimer:The thesis is meant to teach you the scientific approach. Please keep that in mind. We expect you to learn and use scientific methods and to submit a scientifically sound result. Our role is to provide you with everything you need to succeed at that. We expect that you create a piece of work, in which you demonstrate a sound data collection method, an extensive dataset and your analysis of that dataset. A motivated student who works consistently on the thesis throughout the designated period can expect a very good grade. Details on the process and the format that we use to support you in succeeding with your thesis can be found here:
Bachelor: https://bwsyncandshare.kit.edu/s/g7pKCiScRZY5YtH
Master: https://bwsyncandshare.kit.edu/s/MiN8yTkfgeCrnSW

Requirement

Requirements for students

Solid knowledge of Python and machine learning; ideally prior experience with reinforcement learning
Willingness to get up to speed on electricity market fundamentals
Interest in empirical user research
Initiative in recruiting study participants

Faculty departments

Economic & law sciences
Information Engineering

Supervision

Title, first name, last name: Julius Vincent Grams
Organizational unit: KIT WIN - ESIS
Email address: julius.grams@kit.edu
Link to personal homepage/personal page: Website

Application via email

Application documents

Cover letter
Curriculum vitae
Grade transcript

E-Mail Address for application
Senden Sie die oben genannten Bewerbungsunterlagen bitte per Mail an julius.grams@kit.edu

Back