Research
AI systems are moving into places where failure is expensive: healthcare, transport, public safety, critical infrastructure. The research question I care about is what it takes to trust them there — not trust as a slogan, but as an engineering and scientific property that can be specified, built for, tested and argued about.
My approach is to combine learning with reasoning. Machine learning gives systems the ability to handle the messiness of the real world; knowledge representation and reasoning gives them structure, guarantees and the ability to explain themselves. Neither is sufficient alone, and the interesting problems live at the seam between them.
Trustworthy AI
What makes an AI system worth relying on, and how do you demonstrate it? My group works on transparency and explainability, fairness in learned models, synthetic data that preserves utility without carrying forward bias, and the verification and validation of learning-based components in safety-critical systems. This work runs from technical foundations through to the assessment frameworks used in European policy.
Recent directions include intersectional fairness, taxonomies linking fairness and explainability, and — through the RESIST centre — the resilience of AI systems against cyberattack and technical failure.
Autonomous systems
Autonomy is where trustworthy AI stops being abstract. My work here covers motion and mission planning, autonomous 3D exploration of large-scale environments, coordination of multiple agents and vehicles, and unmanned aircraft operating in shared and dynamic airspace. A recurring theme is introspection: systems that model their own competence and know when they are outside it.
This connects to long-running work on stream reasoning — incremental reasoning over continuously arriving, incomplete and noisy data, which is what a robot’s view of the world actually looks like.
Multi-agent systems and decision-making
Coalition structure generation, combinatorial assignment, and multi-objective reinforcement learning: how do you allocate agents to tasks well when the space is enormous, and what happens when there is no single scalar reward that captures what you want? The multi-objective line of work argues that reducing rich objectives to one number discards something essential.
AI education and literacy
If AI capability is going to scale, the bottleneck is people. I work on AI and computational thinking education across the whole range — K-12 curricula and teacher education, university programmes, and professional upskilling. This includes international Delphi studies on what AI literacy should mean, large-scale evaluation of programming in primary education, and comparative policy work on how countries respond to AI skills demand.
The research and the leadership work are the same project seen from two angles: the studies ask what should be taught and how, and the programmes then try to do it at national scale.
Collaboration. My group regularly recruits PhD students, postdocs and research engineers, and I collaborate widely across Europe. If our interests overlap, do get in touch: fredrik.heintz@liu.se.