Expected Minimax Regret Rates for Adaptive LQG Control on Fixed Parameter Sets

Abstract

We study how a fixed model family determines expected minimax regret in adaptive LQG control. The family is compact and affine, with partial observation and unknown noise covariances. We quantify how information from residual covariances and input responses can vanish relative to differences between optimal control laws. This relation yields matching bounds that classify regret as bounded, logarithmic, sublinear polynomial or linear. Polynomial exponents are rational. For real algebraic model and cost data, a terminating procedure determines the regret class and the exact exponent of polynomial growth before data collection. A causal policy attains these rates using model estimation and bounded probes, reselecting model sets from observations without a supplied common stabilizer. For costs with fixed positive definite output and input weights, polynomial exponents are at most one half.