Standard global optimization algorithms, such as Bayesian Optimization (BO) using Gaussian Process (GP) kernels, rely on the assumption of objective function smoothness. However, many real-world discovery tasks involve "regime shifts" or discontinuities where local gradients or covariance structures fundamentally change. We present a "Surprise-Triggered Structural Learning" framework based on Active Inference. By monitoring spikes in Variational Free Energy (VFE)—a formal measure of informational surprise—the proposed Active Inference Engine (AIE) autonomously identifies discontinuities and adapts its internal world-model topology. Benchmarking against standard GP-BO demonstrates critical efficiency gains in reaching target convergence on Heaviside-class discontinuous functions, offering a robust paradigm for optimization in non-stationary and adversarial environments.
Chouhan et al. (Thu,) studied this question.