Simulated annealing with thresheld convergence

Stochastic search techniques for multi-modal search spaces require the ability to balance exploration with exploitation. Exploration is required to find the best region, and exploitation is required to find the best solution (i.e. the local optimum) within this region. Compared to hill climbing which is purely exploitative, simulated annealing probabilistically allows “backward” steps which facilitate exploration. However, the balance between exploration and exploitation in simulated annealing is biased towards exploitation – improving moves are always accepted, so local (greedy) search steps can occur at even the earliest stages of the search process. The purpose of “thresheld convergence” is to have these early-stage local search steps “held” back by a threshold function. It is hypothesized that early local search steps can interfere with the effectiveness of a search technique’s (concurrent) mechanisms for global search. Experiments show that the addition of thresheld convergence to simulated annealing can lead to significant performance improvements in multi-modal search spaces.