Hyper-parameter asynchronous parallel search method based on population evolution

Abstract

The invention provides a hyper-parameter asynchronous parallel search method based on population evolution. The method comprises steps of selecting a member copy weight and a hyper-parameter from a group with excellent performance of the group by utilizing a function; exploring the function to carry out hyper-parameter disturbance; copying the initial hyper-parameter to a new member, and taking the initial hyper-parameter as an initial hyper-parameter; terminating the member with poor performance, and repeating the steps until the hyper-parameter configuration with the optimal performance is found. According to the hyper-parameter asynchronous parallel search algorithm, the advantages of asynchronous parallel calculation and sequence optimization are fused, hyper-parameter search is carried out through utilization and exploration in the thought of the genetic algorithm, the hyper-parameter search time is effectively shortened, and the hyper-parameter search complexity is effectively reduced. In the aspect of hyper-parameter search of deep learning, compared with traditional hyper-parameter search, the performance of the hyper-parameter search is greatly improved, and the hyper-parameter search has certain effectiveness and stability.

Description

【Technical field】 [0001] The invention relates to the technical field of hyperparameter search algorithms, in particular to a hyperparameter asynchronous parallel search method based on population evolution. 【Background technique】 [0002] Deep learning has become a popular machine learning model, which mainly uses multi-layer neural network (Neural networks) training to achieve the desired learning effect. Its motivation is to establish and simulate the neural network of the human brain for analysis and learning, such as image and speech recognition. Deep learning has become a popular machine learning model, which mainly uses multi-layer neural network (Neural networks) training to achieve the desired learning effect. Its motivation is to establish and simulate the neural network of the human brain for analysis and learning, such as image and speech recognition. The concept of deep learning was proposed by Hinton et al. in 2006. It combines low-level features to form more...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to solve the problem that the deep reinforcement learning hyperparameter search time is long and it is difficult to find a better hyperparameter configuration in the prior art, and proposes a hyperparameter asynchronous parallel search method based on population evolution, which can reduce the number of used Computational resources, find better hyperparameters, improve the performance of the algorithm

Images