Artificial Intelligence and Machine Learning have become deeply embedded in our daily lives, often in ways we don’t even notice. Their growing presence has sparked significant interest in applying learning-based methods to control and dynamical systems, especially in the face of uncertainty, noise, and incomplete models. In such settings, learning from expert demonstrations and data-driven approaches can play a crucial role. Interestingly, many learning algorithms themselves can be viewed as dynamical systems, particularly when they involve iterative updates like gradient descent. This perspective opens the door to applying well-established tools from control theory to analyze and improve learning algorithms. In this talk, I will present recent work on developing learning-based solutions for control systems, and explore how viewing learning algorithms as dynamical systems can offer new insights and opportunities for cross-disciplinary innovation.

In digital business, software products must meet diverse and evolving user needs, yet traditional feedback methods like interviews and surveys often prove too slow for agile development cycles. We introduce Active Persona (AP), an AI-based, dynamic user archetype that provides realistic and contextual feedback on demand. By transforming static personas into interactive generative agents, APs enable continuous validation and user-centric experimentation throughout the development process. Our evaluation using a mobile transport app reveals that APs produce insights closely aligned with actual users, identifying the same usability and accessibility issues. These findings demonstrate how APs can accelerate innovation cycles while enhancing user understanding and optimizing resource allocation in digital business environments.

Classical optimal control and estimation techniques can deliver high levels of robust performance, but their practical use in large-scale systems is hindered by dense communication demands. This talk explores how we can implement dense optimal controllers and estimators in a sparse and scalable way. Using ideas from sparse linear algebra, we aim to preserve optimality while drastically reducing the need for all-to-all communication. We illustrate the approach by showing how a simple message passing algorithm can be used to optimally regulate transportation networks.

As software systems grow more complex and development moves faster, ensuring software quality remains both a critical challenge and a moving target. Continuous integration and delivery practices, coupled with the integration of AI components, both in the systems we build and in the tools we use to build them, have transformed traditional quality assurance (QA) approaches. This talk shares insights from our collaborative research efforts to address this challenge.

Autonomous vehicle maneuvering has the potential to significantly improve both traffic safety and transportation efficiency. When the ELLIIT research collaboration on optimal vehicle maneuvers between Dept. Automatic Control (LU) and Div. Vehicular Systems (LiU) was initiated 15 years ago, a key motivation was the increased situational awareness enabled by onboard sensing systems, which opened entirely new possibilities for autonomous control functions. Drawing inspiration from numerical solutions to dynamic optimization problems of relevant maneuvers, led to the development of control principles and real-time architectures for vehicle maneuvering. To address uncertainties in future motion of surrounding vehicles in multi-vehicle scenarios, context awareness is a key feature of the motion prediction and planning. This talk will highlight both the results of this collaboration and aspects of how it was established and developed.