Particle Swarm Optimization for Automated Design of Reversible Residual Neural Networks for Turntable Servo Systems

The reversible residual neural network (RRNN) model is a bidirectional neural network model, which has recently gained attention in the design of various control methods in turntable servo systems due to the practicality of its architecture. However, the current architecture of the RRNN model is mainly designed by manual adjustment relying on prior knowledge, and the inference time of the model need to meet the real-time requirements in the design process. Therefore, an automated neural architecture search method named particle swarm optimization-based architecture search (PSOAS) algorithm for the RRNN model design is proposed in this paper. A new encoding strategy is presented and the possible network architectures in the given search space are encoded while facilitating subsequent updates of positions and velocities during the evolutionary process. To improve the search efficiency of PSOAS, we also develop an early-stopping policy and a memory caching mechanism for reserving the fitness and inference time of the particles during the search process. The experimental results show the superiority of the proposed method against the peer competitors in terms of the prediction performance and search efficiency.