SA-SGA: Simulated Annealing Optimization and Stochastic Gradient Ascent Reinforcement Learning for Feature Selection

Selecting appropriate features from the dataset can increase the model’s efficiency and reduce its complexity. The problem of feature selection is NP-hard, so solving it accurately can be time-consuming. To reach the best solution, metaheuristic algorithms might be the best choice. For solving the optimization problem, simulation annealing (SA) uses a probabilistic approach. It is possible to get trapped in local minima if the temperature is not lowered gradually during the annealing process. It is also difficult to find a suitable neighbor in SA. In this paper, reinforcement learning is used to address the weaknesses of SA. As a result of allocating rewards for the selection of features, the feature selection problem forms a policy for moving to a new neighbor. To adjust the time spent at each temperature, the average reward received is used. Using the stochastic gradient ascent algorithm, the multi-armed bandit plays according to the degree of priority. Each feature is considered an arm and selected as an arm pull to perform the simulation. The proposed algorithm is named simulated annealing—stochastic gradient ascent. Finally, different algorithms are compared, including the SA, the grasshopper optimization algorithm (GOA), the enhanced GOA (EGOA), the grey wolf optimization (GWO), the nondominated sorting genetic algorithm-II, and the nondominated ranking genetic algorithm (NRGA). The proposed algorithm was found to improve the objective function by 24.6%, 14.7%, 2.46%, 11.4%, 7.8%, and 6.9% compared to SA, GOA, EGOA, GWO, NSGA-II, and NRGA, respectively.