A non-revisiting simulated annealing algorithm

Abstract

In this article, a non-revisiting simulated annealing algorithm (NrSA) is proposed. NrSA is an integration of the non-revisiting scheme and standard simulated annealing (SA). It guarantees that every generated neighbor must not be visited before. This property leads to reduction on the computation cost on evaluating time consuming and expensive objective functions such as surface registration, optimized design and energy management of heating, ventilating and air conditioning systems. Meanwhile, the prevention on function re-evaluation also speeds up the convergence. Furthermore, due to the nature of the non-revisiting scheme, the returned non-revisited solutions from the scheme can be treated as self-adaptive solutions, such that no parametric neighbor picking scheme is involved in NrSA. Thus NrSA can be identified as a parameter-less SA. The simulation results show that NrSA is superior to adaptive SA (ASA) on both uni-modal and multi-modal functions with dimension up to 40. We also illustrate that the overhead and archive size of NrSA are insignificant, so it is practical for real world applications.