SPGD: Search Party Gradient Descent Algorithm, a Simple Gradient-Based Parallel Algorithm for Bound-Constrained Optimization

Abstract

Nature-inspired metaheuristic algorithms remain a strong trend in optimization. Human-inspired optimization algorithms should be more intuitive and relatable. This paper proposes a novel optimization algorithm inspired by a human search party. We hypothesize the behavioral model of a search party searching for a treasure. Motivated by the search party’s behavior, we abstract the “Divide, Conquer, Assemble” (DCA) approach. The DCA approach allows us to parallelize the traditional gradient descent algorithm in a strikingly simple manner. Essentially, multiple gradient descent instances with different learning rates are run parallelly, periodically sharing information. We call it the search party gradient descent (SPGD) algorithm. Experiments performed on a diverse set of classical benchmark functions show that our algorithm is good at optimizing. We believe our algorithm’s apparent lack of complexity will equip researchers to solve problems efficiently. We compare the proposed algorithm with SciPy’s optimize library and it is found to be competent with it.