Implementation Of Task Structure Utilization In Four Machine Learning Tasks

Machine learning aims at studying how to improve the performance of the system by using experience.In a computer system,"experience" often exists in the form of "data".Therefore,machine learning mainly focuses on the algorithm of learning"model" from data of tasks.In the real world,there are many different forms and structures of data,depending on their source and the goals of machine learning tasks,and thus lead to different tasks.For example,Wikipedia entries may have one or more labels,and the dimension of labels is tens of thousands.If we want to classify each entry accurately,we need to establish a high-dimensional multi-label learning task to learn it.For the image data generated by users on the Internet,although it is unlabeled,it still contains a lot of user behavior information.In order to find potential information in high-dimensional unlabeled data such as images,we need to establish high-dimensional unlabeled learning tasks.In addition,in some machine learning problems,such as reinforcement learning problems,data is generated in the process of interaction with the training environment.In order to effectively utilize the hierarchical data generated in complex environments,we need to establish hierarchical reinforcement learning tasks.And for the empirical data in a batch of similar tasks,we need to establish policy experience reuse learning tasks to improve the performance of reinforcement learning.1.Traditional multi-label learning algorithms face the problem of dimensionality disaster in high-dimensional multi-label learning tasks.In this paper,we analyze and make use of the relevance of labels,and propose a multi-label classification algorithm BILC based on derivative-free optimization.Through derivative-free optimization,we compress high-dimensional labels to low-dimensional binary subspace.And thus we can use classifiers instead of regressors to predict in lowdimensional sub-space,which leads to lower structure risk.BILC greatly improves the effectiveness of multi-label learning algorithm on high-dimensional multi-label data.2.In high-dimensional unlabeled learning tasks,the traditional mixture model clustering algorithm is difficult to capture the distribution of complex data.This paper combines the mixture models based clustering and the Generative Adversarial Network(GAN).And then proposes the Mixture of GAN Models(GANMM),which leads to end-to-end clustering on raw image data.We also propose ?-EM procedure to learn GANMM.And finally GANMM shows good performance on highdimensional unlabeled data clustering.3.For hierarchical reinforcement learning tasks on complex environments,current hierarchical reinforcement learning approach only use these data to learn multilevel policies,which is difficult to decide to high-level decision frequency.In this paper,we propose Temporal-Adaptive Hierarchical Reinforcement Learning,called TEMPLE.It not only abstracts the hierarchies of data,but also utilizes the temporal information in data to adjust high-level decision frequency by its temporal structure.And TEMPLE outperforms other approaches on many tasks.4.In policy experience reuse learning tasks,current methods only extract and reuse empirical data from a certain granularity,which will lose some accuracy and limits the performance of policy experience reusing.In this paper,we propose a multigranularity and multi-level experience extraction and reusing algorithm,PRR.And we also propose a top-down training approach for PRR.PRR shows good performance on policy experience reuse tasks.