Hyperparameter Search Based On Population Evolution And Its Application In Manipulator Grasping Imitation Learning

With the rapid development of hardware computing capabilities,especially large-scale distributed parallel computing,the field of machine learning has made great progress.Under the premise that the training data is sufficient,the problem of hyperparameter configuration of machine learning algorithms is the key to achieving good results.Hyperparameters are parameters that need to be selected before a machine learning algorithm runs,such as those in deep learning algorithms that control the learning rate of the neural network learning speed.The purpose of hyperparameter search is to select a good set of hyperparameters for an application's algorithm,so that the performance of this algorithm is optimal.In the previous machine learning hyperparameter selection problem,researchers in the field generally selected the hyperparameters manually based on personal experience.With the exponential increase in data scale,although the rapid progress of large-scale computing acceleration equipment has made deep learning,especially deep reinforcement learning algorithms,show strong advantages in processing massive data such as pictures,it is still a problem in the selection of hyperparameters.Unsolved puzzle.In recent years,with the continuous increase of deep learning,especially deep reinforcement learning models,the training cost,that is,the search space of hyperparameters,has also continuously increased.However,most traditional hyperparameter search algorithms are based on sequential execution of training,and often need to wait It may take weeks or even months to find a better hyperparameter configuration.In order to solve the problem of long search time for deep reinforcement learning and difficult to find a better hyperparameter configuration,this thesis proposes a new hyperparameter search algorithm based on population evolution Hyperparametric Asynchronous Parallel Search(PEHS).The algorithm combines evolutionary algorithm ideas and uses a fixed resource budget to asynchronously search the population model and its hyperparameters in order to improve the performance of the algorithm.The thesis designs and implements a parametric search running on the Ray parallel distributed framework Algorithms,experiments show that the performance of asynchronous parallel search of hyperparameters based on population evolution on a parallel framework is better than that of traditional hyperparameter search algorithms,and the performance is stable.Manipulator is the most widely used mechanical device in the field of robotics,and its behavior acquisition research is an important aspect of robot kinematics research.Teach-to-learn is a fast and efficient way to learn.By learning the teaching behavior,the robotic arm can quickly acquire motor skills,thereby simplifying complex movement planning and improving learning efficiency.This thesis focuses on the study of robotic arms to obtain motion behavior through teaching.In this thesis,a convolutional neural network model(CNN)is designed,which uses superparametric asynchronous parallel search algorithm(PEHS)based on population evolution to conduct research on KUKA manipulator based on imitation learning through supervised learning.This method can make use of limited hardware resources to quickly evaluate the network architecture,and at the same time ensure that each network trained in parallel selects good hyperparameters that adapt to the current model during the optimization process,making the model more accurate and robust.The experiments show that the combined PEHS algorithm can effectively improve the experimental accuracy by about 6% compared with the supervised learning(data aggregation)method,and the stability of the algorithm is stronger.