LSTM V-Network Swarm Optimizer(LVNSO): A New Meta-Heuristic Based on Machine Learning Methods

Traditional meta-heuristics have good performance on solving black box problems with flexibility, derivation-free mechanism and local optima avoidance. How-ever, due to the simplicity of most models, traditional meta-heuristics often don't have high stability and reliability on complex continuous problems. This work proposes a new meta-heuristic called LSTM V-Network Swarm Optimizer (LVNSO) inspired by machine learning models and methods. The LVNSO algorithm has a basic structure of a swarm with co-evolutionary of leader performance and LSTM V-Network in iteration, with ∊-greedy exploration in reinforcement learning to avoid local optima. Additionally, resetting the network is aim to recover the sensitivity of the net, perform local optimization and improve the accuracy of the result. The algorithm is tested by 23 CEC 2005 benchmark functions and is verified by comparative study with several outstanding algorithms. The results show that the LVNSO is able to give much stabler and more accurate result than other algorithms overall under the condition of different parameters adapted to different functions.