AI4REALNET is a Horizon Europe project on the use of AI in critical network infrastructure. It was selected as one of four projects from 114 submissions, and it brings together grid and transport operators, research institutes and companies building the systems. enliteAI is a partner, and we co-authored the project’s position paper.
The paper is an attempt to be specific about something the field usually discusses in the abstract: what AI in network operations would actually consist of.
Four things the paper covers
Where the potential is. Not “AI will transform infrastructure”, but which operational decisions in electricity, railway and air traffic networks are the ones where an algorithm has an advantage over an expert under time pressure — and which are not.
How humans and AI share the work. Critical networks are operated by people whose expertise is increasingly augmented by control and supervision software at varying levels of automation. The paper looks at what makes that cooperation improve over time instead of degrading, and at the technical and ethical questions it raises.
The environments needed to train agents. This is the part outsiders underestimate. An agent cannot be trained on a live grid. Someone has to build a simulation faithful enough that a policy learned inside it survives contact with the real network — which means the physics, the contingencies and the operational constraints all have to be in there.
Where the research needs to go. Advanced learning techniques, and co-learning models in which the human and the system adapt to each other rather than the human simply supervising a fixed policy.
Why we work in consortia
Almost everything we can prove about our own methods rests on infrastructure someone else built and shared: grid simulation environments, benchmark networks, competition frameworks. Publishing back into that pool is not altruism. It is how a small research team gets access to problems at the scale of a real network, and how claims made in this field become checkable by people other than the company making them.
Authors
From enliteAI: Anton Fuxjäger, Alberto Castagna, Yassine El Manyari, Stefan Zahlner and Marcel Wasserer, with the AI4REALNET consortium including Antoine Marot, Jan Viebahn and Ricardo Bessa.
