Master Thesis, 30 hp: Valid and Reachable Counterfactual Explanations for Reinforcement-Learning Policies in Pursuit–Evasion

Are you a student eager to apply your theoretical knowledge and fresh perspectives to real-world challenges? At Saab, we believe that innovation thrives on new ideas, and your master thesis project could be the spark that ignites our next technological breakthrough.

Your role

We recognize the immense value that students bring to our company. Your academic rigor, combined with your enthusiasm for cutting-edge technology, allows you to approach problems with a unique and insightful lens. At Saab, you'll have the opportunity to collaborate with experienced engineers and specialists, gaining invaluable practical experience while making a tangible contribution to our growth and development.

Background
Without interpretable rationale or "what-if" counterfactuals, machine learning engineers/developers spend excessive time manually inspecting logs for debugging, leading to “improper” Reinforcement Learning (RL) policy patches that may or not improve the policy’s performance. Understanding of a RL agent’s decision outcomes improves the developer’s time to debug and model steer the RL policy in training that achieves a good performance in deployment.

Description of the Thesis Project

This thesis will investigate to what extent do minimal counterfactual state changes provide faithful, feasible, and stable explanations of an RL agent’s decisions in a sequential pursuit-evasion environment? The student will implement a baseline method that identifies minimal state changes required to alter an agent’s action and compare it with a dynamics-constrained method that generates only feasible and reachable counterfactuals. The counterfactuals explanations can be quantitatively evaluated for action validity, reachability, stability, intervention faithfulness, and trajectory-level consequences. Furthermore, a simple user interface to visualize the counterfactuals will be developed and evaluated through user studies. The expected outcome is both an XRL demonstrator and guidelines on when counterfactual explanations can and cannot support understanding of autonomous decision-making.

The master thesis student will be part of the research project about unmanned collaborative aircraft systems. Furthermore, the master thesis student will engage with the organizational unit called Planning, Optimization and Reinforcement Learning through working relation (joint presentations and informal discussions) with the unit’s RL evaluation and debugging team. The resulting XRL tool will be added to the toolbox and made available to the unit’s evaluation and debugging capabilities of advanced AI techniques.

Your profile

The project is suitable for students with an interest in reinforcement learning, eXplainable AI and interface design using visualization techniques. You are in the final stage of civilingenjör programmes – Computer Science and Engineering, Systems and Controls, Autonomous Systems, Robotics and AI and Machine Learning profile or equivalent and are about to begin a 30 hp thesis project. You have experience with PyTorch, TensorFlow or JAX. You have a good understanding of MDPs, policies, value functions and deep RL. You have the ability to work with an existing research codebase using Git. You display basic knowledge in statistics, e.g., confidence intervals, multiple random seeds, hypothesis testing and interpreting variance. Finally, you are interested in analysing why an agent behaves as it does.

We provide the support and guidance you need to translate your theoretical knowledge into practical solutions. Join us and become a driving force behind Saab's technological advancements!

This position requires that you pass a security vetting based on the current regulations around/of security protection. For positions requiring security clearance additional obligations on citizenship may apply.

Kindly observe that this is an ongoing recruitment process and that the position might be filled before the closing date of the advertisement.

Contact information

Emil Karlsson, Head of Planning, Optimization and Reinforcement Learning

073-4180368

emil.karlsson1@saabgroup.com

Jenni Reuben, Master Thesis Industrial Supervisor

073-446 8110

jenni.reuben@saabgroup.com

What you will be a part of

Explore a wealth of possibilities. Take on challenges, create smart inventions, and grow beyond. This is a place for curious minds, brave pioneers, and everyone in between. Together, we achieve the extraordinary, each bringing our unique perspectives. Your part matters.

Saab is a leading defense and security company with an enduring purpose, to help nations keep their people and society safe. Empowered by its 28,000 talented people, Saab constantly pushes the boundaries of technology to create a safer and more sustainable world. Saab designs, manufactures and maintains advanced systems in aeronautics, weapons, command and control, sensors and underwater systems. Saab is headquartered in Sweden. It has major operations all over the world and is part of the domestic defense capability of several nations. Read more about us here.