Causal Models for Decision Systems: An Interview with Matteo Ceriscioli

The AI community is increasingly focused on building systems that remain reliable when their operating conditions change. Distribution shifts—whether due to seasonal demand, sensor drift, or policy updates—can degrade model performance dramatically. Ceriscioli’s work demonstrates that an agent’s ability to adapt is mathematically equivalent to possessing a causal model of its environment, turning robustness into a measurable property of causal knowledge. This insight reframes robustness research, encouraging developers to embed causal inference directly into learning pipelines rather than treating it as an after‑thought.

Building on that foundation, Ceriscioli introduced causal partially observable Markov decision processes (POMDPs) to let robots reason about unknown interventions. By maintaining a belief over both state and potential environmental changes, a robot can autonomously detect and compensate for shifts, such as altered terrain or sensor noise, without human re‑training. The same framework supports transfer learning: a well‑trained agent can export its causal representation as a prior for a new agent operating in a related domain, accelerating training for tasks like drone navigation after a rover has learned the underlying terrain dynamics.

Despite these advances, practical causal discovery remains a bottleneck. Real‑world data are often incomplete, and existing algorithms struggle with missing values that bias causal estimates. Ceriscioli’s current efforts target scalable methods that extract causal structures from both adaptable agents and imperfect observational datasets. Success would enable industries—from telecom churn mitigation, where he previously applied causal models at Vodafone, to autonomous logistics—to deploy AI that not only predicts outcomes but also understands the levers that drive them, fostering safer, more trustworthy systems.