Reinforced Optimization: An Objective-Based Optimization Algorithm for Reinforcement Learning

Reinforcement learning (RL) is a machine learning paradigm that enables systems to learn from experience within uncertain environments. However, RL is typically suited to goal-oriented problems rather than objective-driven tasks. To address this limitation, this paper proposes a new optimization algorithm, called reinforced optimization (RO). This algorithm leverages an RL framework and reformulates it from a goal-based setting into an objective-based formulation, enabling its application to optimization problems. The algorithm preserves essential RL elements such as rewards and penalties, as well as exploration and exploitation mechanisms. It utilizes sequences of rewards to represent problem states and updates these states through directional steps. The effectiveness of each action with respect to the objective function is evaluated: successful actions lead to increased rewards, while unsuccessful actions result in decreased rewards. This iterative process continues for a predefined number of iterations, aiming to achieve convergence towards a global optimum in optimization tasks. To evaluate the RO algorithm, experiments were conducted on three objective-based problems: centroid optimization, traveling salesman problem, and vehicle routing problem. The algorithm’s performance was compared to other optimization methods: genetic algorithm, ant colony optimization, simulated annealing, and particle swarm optimization. The results indicate that the RO algorithm exhibits behavior analogous to gradient descent during convergence and demonstrates a favorable trade-off between accuracy and execution time, making it a viable alternative for both continuous and combinatorial optimization problems.