Controller Design for the HDD Benchmark Problem using RNN-based Reinforcement Learning

With the rapid spread of the cloud server market, hard disk drives (HDDs) are required to achieve higher truck following speeds and storage capacities. To meet these demands, it is necessary to improve the accuracy of positioning control of the magnetic head that writes data to the disk. In this paper, we propose a controller designed by recurrent neural network (RNN-) based reinforcement learning for the “benchmark problem of designing a control system for a hard disk drive with two-stage actuators” and evaluate the performance of the RNN-based controller. We also propose a method to convert an RNN-based controller to a state-space linear controller, which guarantees stability and enables performance analysis. This contributes to eliminating the black box drawback of reinforcement learning. Performance verification on the benchmark problem shows that the 3σ value of the head position of the track pitch of the RNN-based controller shows an accuracy improvement of approximately 1.8-fold compared to the reference (example) controller. The sensitivity function of the RNN-based controller shows superior disturbance attenuation especially for the disturbance due to external vibration caused by air-cooling fans and other HDDs in a storage box. The article describes the structure of the proposed RNN-based controller and presents an evaluation of the benchmark results(1).