A Reinforcement Learning-Based Bi-Population Nutcracker Optimizer for Global Optimization.

The nutcracker optimizer algorithm (NOA) is a metaheuristic method proposed in recent years. This algorithm simulates the behavior of nutcrackers searching and storing food in nature to solve the optimization problem. However, the traditional NOA struggles to balance global exploration and local exploitation effectively, making it prone to getting trapped in local optima when solving complex problems. To address these shortcomings, this study proposes a reinforcement learning-based bi-population nutcracker optimizer algorithm called RLNOA. In the RLNOA, a bi-population mechanism is introduced to better balance global and local optimization capabilities. At the beginning of each iteration, the raw population is divided into an exploration sub-population and an exploitation sub-population based on the fitness value of each individual. The exploration sub-population is composed of individuals with poor fitness values. An improved foraging strategy based on random opposition-based learning is designed as the update method for the exploration sub-population to enhance diversity. Meanwhile, Q-learning serves as an adaptive selector for exploitation strategies, enabling optimal adjustment of the exploitation sub-population's behavior across various problems. The performance of the RLNOA is evaluated using the CEC-2014, CEC-2017, and CEC-2020 benchmark function sets, and it is compared against nine state-of-the-art metaheuristic algorithms. Experimental results demonstrate the superior performance of the proposed algorithm.