Abstract
Metaheuristics optimization algorithms have established themselves as a new cornerstone in the area of complex, nonlinear, and high-dimensional problems of science and engineering. Most existing algorithms, however, still do not manage to balance exploration and exploitation and still experience stagnation and premature convergence. Though there have been many developments, most of the available algorithms are not yet optimized in terms of premature convergence and adaptive decision-making ability. This paper presents two smart reinforcement-based schemes, one Adaptive Intelligent Reinforced Walrus-Gazelle Optimizer (AIRE_WGO), and another Dual-Mode Adaptive Reinforced Switching Walrus-Gazelle Optimizer (DMARS_WGO). The suggested AIRE_WGO steps up the classical Walrus gazelle hybrid, where it implements Q-learning-based control of the agent behavior, adjusts its parameters in response to the current environment, and considers diversity-informed mutations in order to dynamically switch between global exploration and local exploitation. However, DMARS_WGO incorporates a dual-agent reinforcement framework that combines tabular Q-learning and deep Q-network (DQN) reasoning. When interacting with diverse populations, DMARS_WGO autonomously switches between its learning dominance as a result of real-time feedback about the population structure, the rate of improvement, and the level of stagnation. Moreover, the cross-agent knowledge-sharing process provides a two-way experience transfer between Q-learning and DQN modules, which strengthens cooperative intelligence and stability. Extensive tests of CEC2017 and CEC2022 benchmark suites and six engineering design problems prove that the proposed algorithms, especially DMARS_WGO, are more successful compared to nine recent state-of-the-art optimizers including Golden Jackal Optimization (GJO), Osprey Optimization Algorithm (OOA), Pelican Optimization Algorithm (POA), Rat Swarm Optimizer (RSO), Smell Agent Optimization (SAO), Channa argus Optimizer (CAO), Rüppell's Fox Optimizer (RFO), Mantis Shrimp Optimization Algorithm (MShOA), and Adaptive L-SHADE (ALSHADE). On CEC2017, DMARS_WGO achieved first rank in 26 out of 29 benchmark functions and obtained the best overall Friedman mean rank. On CEC2022, it ranked first in 8 out of 12 functions and achieved the best overall ranking among all compared algorithms. Similarly, across the six engineering design problems, DMARS_WGO secured first position in 4 out of 6 cases and achieved the best overall ranking performance. Wilcoxon Signed-Rank and Friedman mean-rank statistical tests prove that DMARS_WGO is significantly superior. Its ability to smartly self-adapt its search dynamics is pointed out in the results, and makes it a strong and robust optimizer of real-world engineering problems.