Adaptation and linear-response theory
Toyonori Munakata, Satoshi Oyama · Physical review. E, Statistical physics, plasmas, fluids, and related interdisciplinary topics · 1996
Adaptation, i.e., slow time variation of control parameters to achieve better performance, which is usually estimated by some object or cost function, is investigated based on linear-response theory. Adaptive Monte Carlo methods such as entropy (or multicanonical) sampling and nonsupervised Hebbian learning are derived from the criterion that the cost (or gain) should be expressed in terms of (generalized) autocorrelation response functions. Our approach also gives an approximate expression for the cooling rate in simulated annealing in terms of heat capacity and energy diffusion. \textcopyright{} 1996 The American Physical Society.