Minimax adaptive control for finite sets of positive linear systems

What the math gives to ML

The paper provides a constructive minimax adaptive-control pattern for systems whose dynamics belong to a finite set: maintain candidate plants, let an adversary choose the worst one-step successor, and select a certainty-equivalent feedback law using a Bellman-style cost certificate. The transferable asset is not positivity itself, but the combination of finite-model uncertainty, state-dependent action constraints, and an explicit worst-case successor that avoids requiring an initially stabilizing policy. This suggests a robust recurrent or state-space neural module with a small bank of candidate transition matrices and online minimax mode selection. The first test should measure whether this improves hidden-state stability and robustness to temporal distribution shifts at comparable parameter count and training cost.