Neural networks are attractive for controlling systems whose dynamics are too complex to model from first principles. But when a network is trained offline and then placed inside a feedback loop, its own outputs change the very distribution of states it was trained on. Vittorio Lippi's paper, posted on arXiv, formalizes this failure mode and proposes a practical fix.

Lippi introduces a taxonomy of three sequential errors: the initial closed-loop identification error, the deployment-induced action error, and the continuous tracking error. Using discrete-time sensitivity analysis, he shows that offline-learned parameters locally guarantee a trajectory divergence once deployed, because the loss landscape that shaped the parameters no longer matches the closed-loop data distribution.

To counter this, the paper proposes time-scale separated online adaptation, in which the network continues to learn at a slower timescale than the control loop operates. In ablation studies on a nonlinear Duffing oscillator, bounded continuous learning kept trajectories on track, while static deployments diverged. The author argues that practitioners building safety-critical machine-learning control systems should plan for online adaptation rather than treating deployment as a one-time event.