IcsP 2020 IOP Publishing Tournal of Physies: Conference Series 11544 2020) 012003 doi 10-1088/1742-6596/1544/1/012003 Research on Machine Learning and Its Algorithms and Development Wei Jin Northwestern Polytechnical University Ming De College, Xian, Shaanxi, China Abstract: This article antlyzes the basic classification of machine learning, including supervised learning, unsupervised learning, and reinforcement learning. It combines analysis fon common algorithms in machine learning, such as decision tree algorithm, random forest algorithm, artificial neural network algorithm, SVM algorithm, Boosting and Bagging algorithm, BP algorithm, Through the development of theoretical systems, further improvement of autonomous learning capabilities, the integration of multiple digital technologies, and the promotion of personalized custom serviees, the purpose is to improve people's awareness of machine learning and accelerate the speed of popularization of machine learning. 1. Introduction With the rapid development of science and technology, artificial intelligence has also ushered in new development opportunities. Machine technology based on computer technology incorporates multidisciplinary theoretical knowledge, such as statistics and algorithm complexity, which further strengthens the functional attributes of artificial imelligence. By doing a reasonable analysis of ‘machine learning algorithms, it can provide direction reference for subsequent machine learning development, thereby improving the applicability of machine learning algorithms and providing more convenience for the economic development of the industry. 2. Basic Classification of Machine Learning 2.1. Supervised Learning In the process of machine learning, supervised learning belongs to a relatively basic leaning method. This learning method refers to the establishment of corresponding learning goals by people before learning. During the initial training of the machine, the machine relies on information technology to learn the needs of learning. In order to collect basic data information, we are supposed to gradually complete the required learning content in a supervised environment. Compared with other learning ‘methods, supervised learning can fully stimulate the generalized learning potential of the machine itself. After completing the system learning, it can help people to solve some classification or regression problems, which is highly systematic. Currently, the classic learning methods commonly used include BN, SVN, KNN, etc. Because the entire leaming process has purpose, the machine learning process presents a certain regularity, and the learning content is more systematic [1]. 2.2. Unsupervised Learning Corresponding 10 supervised leaming is unsupervised Icing, The so-called unsupervised leaming rcens thatthe machine dacs not mark the content in a certain diootion during the entire Ieaming coe om i way he ed dt et th Cte Cams ttn 3 ay er iin ea Jof this work must maintain atribution to the author(s) and the ttle of the work, journal citation and DOL. Frieden OP Psshng 7IcsP 2020 IOP Publishing Tournal of Physics: Conference Series 11544 2020) 012003 doi 10-1088/1742-6596/1544/1/012003 process, but rely on the machine itself to complete the analysis of data information, In practice, the ‘operation method is to let the machine learn the basic concepts and content, and then give the machine enough freedom to complete a series of content learning, including concepts and content similar to the basic principles, such as tree roots, In general, the continuous improvement of learning in stages has increased the breadth of machine learning content. At present, unsupervised learning includes algorithms such as deep belief networks and autoencoders. Such situations are conducive to the solution of clustering problems and have good applications in the development of many industries [2] 2.3. Reinforcement Learning In addition to supervised learning and unsupervised learning, there are also application methods of reinforcement learning in machine learning. The so-called reinforcement learning is the systematic learning of a certain content, In the specific application process, the data collected in the previous period will be used. It organizes and processes the feedback information of a certain part to form a closed loop of data processing. On the whole, reinforcement learning is a type of learning method that expands data collection based on statistics and dynamic learning. Such methods are mainly used to solve the control problem of robots. Its representative learning methods include Q-learning algorithm and Temporal difference learning algoritlam. 3. Analysis of Commonly Used Algorithms for Machine Learning 31, Decision Tree Algorithm Among the commonly used algorithms for machine learning, the decision tree algorithm belongs to the classic algorithm content. Its working principle is that when processing data information, it starts from the root node of the collection instance and reaches the position where the nodes meet to make it complete. Scientific division of practical examples. In order to facilitate the analysis of data information, the decision number algorithm will continue to split branches, and at the same time, the branches will be trimmed to improve the integrity of the data content [3]. From the point of view of calculation, the algorithm belongs to the top-down algorithm. During the content analysis process, the content of the node is analyzed for the optimal attributes, and then the node is expanded to more than ‘two based on the node. This way, you can get comprehensive data information of the split, and the branching method like a tree can also increase the number of samples that can be analyzed, and at the same time determine the content that contains the most samples in the classification according to the sample number statistics. For example, when analyzing data, you can name the decision tree with a large amount of data information as the larger tree A, and set the upper limit of branch splitting. If the ‘upper limit is set to 5, the larger tree A is in the classification afer reaching the value of 5, it will stop continuing to split, and at the same time use the pruning strategy to process the larger tree model, 0 sto refine the data and improve the seientificity of the data analysis results 3.2, Random Forest Algorithm Similar to the decision tree algorithm, in the process of data calculation, the random forest algorithm, ccan be used for further processing. The random forest algorithm will play a good role in controlling unreasonable data in the process of actual use. Thereby effectively improving the scientificity of the data split results and the accuracy of the data analysis results. At the same time, in the process of data analysis, multiple sets of classification trees will be created at the same time, and then the unified algorithm will be used for regression processing. Assuming the decision tree is an independent set ai (i .2.3 ...), then the random forest is the total set A, where A= {al, a2, 23... an}, where 1 Each set remains independent, and the distribution is a state of random distribution, When evaluating the classification data information, it will be selected by means of voting, The classification with the highest number of votes in the voting will output the vector value xi, and then the vector content will be classified to calculate the average value of different score states and provide data reference for the final judgment [4].IcsP 2020 IOP Publishing Tournal of Physics: Conference Series 11544 2020) 012003 doi 10-1088/1742-6596/1544/1/012003 3.3. Artificial Neural Network Algorithm The so-called artificial neural network refers to imitating the process of human information transmission, classifying different data into one neuron, and connecting the data neurons with the help of the Internet to achieve complex memory activities, However, the artificial neural network algorithm, is based on this unfolding data analysis process. Among the delineated neurons, each digital unit has a high degree of authenticity, and the data can complete the process of external output. I's just like the human body moves forward, stops, and runs. In the artificial neural network algorithm, the data information presented has a varicty of application characteristics, and the corresponding analysis process can be completed according to actual needs. At present, commonly used artificial neural networks include multilayer forward neural networks MLEN, self-organizing neural networks, SOM, and ART [5]. In order to facilitate the analysis and calculation of the data, we can set the weighting coefficient in advance and then set the output threshold. After the calculated sum exceeds this value, a certain value is output to the outside, thereby improving the orderliness of the entire numerical analysis process, 3.4. SVM Algorithm In the process of machine learning, the SVM algorithm also belongs to the commonly used algorithm content. In the specific application process, the algorithm mainly relies on the vector machine method to complete the established data analysis work. At the same time, the SVM algorithm will use the automatic support of the SVM (o analyze the data information to be processed, so as (0 optimize the data information. In order to improve the scientificty of the final data analysis results, in the actual analysis process, multiple sets of analysis samples need to be collected to determine the sample data of the boundary value. For example, assuming that the data information to be processed is H (@), when processing it, first, the data information is processed centrally with the help of SVM technology so that it can be completely dispersed. Secondly, the boundary of the H (d) plane is determined from the ‘maximum distance of the entire plane. Finally, the vector content of the Hi (d) plane is analyzed to obtain the output vector, which improves the accuracy of data processing. 4.5. Boosting and Bagging Algorithms Boosting algorithm as a new type of machine algorithm content, its biggest application advantage is that it can complete the accurate processing of data information and improve the accuracy of the final processing result, In practice, the function prediction system will be built with the help of Boosting algorithm, and the system content will be continuously optimized with the help of reinforcement learning mode, thereby speeding up the processing of data information, AdaBoost is a relatively basic application in the Boosting algorithm. At the same time, AdaBoost is also an important guarantee for the expansion of the Boosting algorithm, The Bagging algorithm has a high similarity in the data processing process. In actual application, the difference is that the Bagging algorithm randomly selects, the training set. And during the calculation of the function model, the Bagging algorithm does not analyze the weight content, and we need to continuously optimize the data model with the help of training to improve the accuracy of the data analysis results. 3.6, BP Algorithm ‘The BP algorithm belongs to supervised learning. The basic principle of the algorithm is shown in Figure 1, The figure shows a shallow forward neural network computing model, which includes an input layer, a hidden layer, and an oulput layer. A large number of neurons are connected to each other as network nodes. . Each neuron processes the connection strength signals as network weights through an excitation function. By adjusting these connection strengths, the pattern information contained in the input data is mapped to the output layer.IcsP 2020 IOP Publishing Tournal of Physics: Conference Series 11544 2020) 012003 doi 10.1088/1742-6596/1544/1/012003 Hidden Layer " Expected Output ¥ layer ty layer L, Figure 1 Basic Principles of Algorithm Application As can be seen from the above figure, the direction of the information flow of forward propagation. is input layer > hidden layer > output layer, and its mathematical model is: bu. s(x) =£C9) Haxt 8). Co . Where Wi and b are their weights and bias parameters, f (W, b; x): R- — Ris called the excitation function, and sig-moid can be selected in practical applications, Tan, ReLU and other functions or their variants, hv. (X) are the network output values. In practical applications the BP algorithm can be implemented by the steepest descent method, Newton method and its improved algorithm, quasi-Newrton method and its correction algoritim, ct. At presen, the L-BFGS algorithm is most widely used, and non-precise line search methods are often used to complete the optimization. This method follows Wolfe's criterion and Armijo's criterion, whi guarantees the balance between the decline of the cost function and the convergence of the iterative Sequence 4. Research on Machine Learning Development 4.1. Theoretical System Continues to Mature In the future development process, the mechanical theory system will also be further optimized, and its content branches and coverage will also be expanded, In the initial formulation process of machine learning content, its content is mainly applicable to some automation industries, and the content of the ‘entire theoretical system has not been completely sound. In practical application, the content of its theoretical system is not applicable in some fields. In response to such situations, the next stage of machine learning theory will be continuously strengthened, and the degree of refinement of the content will also be strengthened, which provides convenient conditions for the subsequent promotion of machine learning, 4.2, Autonomous Learning Ability is Further Improved At present, many enterprises in China have realized the development model of automation, and intelligence is the focus of the next stage of development. In the context of the rapid development of