東華大學圖書館 |

Machine Learning-Based Decision Making in Autonomous Systems.

Record Type:	Electronic resources : Monograph/item
Title/Author:	Machine Learning-Based Decision Making in Autonomous Systems./
Author:	Ghazanfari, Behzad.
Published:	Ann Arbor : ProQuest Dissertations & Theses, : 2020,
Description:	266 p.
Notes:	Source: Dissertations Abstracts International, Volume: 82-06, Section: B.
Contained By:	Dissertations Abstracts International82-06B.
Subject:	Computer science. -
Online resource:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28256632
ISBN:	9798557010252

Machine Learning-Based Decision Making in Autonomous Systems.
Ghazanfari, Behzad.

Machine Learning-Based Decision Making in Autonomous Systems. - Ann Arbor : ProQuest Dissertations & Theses, 2020 - 266 p.

Source: Dissertations Abstracts International, Volume: 82-06, Section: B.

Thesis (Ph.D.)--Northern Arizona University, 2020.

This item is not available from ProQuest Dissertations & Theses.

Machine learning-based decision making is a statistical process of selecting different actions or assignments depending on input data. Autonomous decision making in machine learning (ML) is expected to handle different conditions such as non-linear environment, uncertainties in observations, and complexities of real-world domains in efficient and robust ways. Machine learning methods are categorized into reinforcement learning (RL), unsupervised learning approaches, and supervised learning methods. Clustering and classification are paradigms of unsupervised learning and supervised learning correspondingly. RL, classification, and clustering are applied depending on the type of problem. RL is considered as a learning approach for sequential decision-making problems. RL learns under uncertainty to maximize accumulated reward by trial and error, in which the reward signals may be received by delay. Classification and clustering are used for one-step decision-making problems. Classification and clustering assign instances to groups. There are labels in the classification problems while there are no labels in clustering problems. There are different types of static datasets such as images, documents, or classical ones, and non-stationary datasets like time-series.There are several issues for applying ML approaches in real domains: 1) expert knowledge is required to train approaches, which is not always cost-effective or feasible, 2) lack of generalizable approaches across multiple domains of ML, 3) inefficiency in handling complexities and non-linearity of patterns while there are considerable normal diversities in instances, 4) inefficiency in handling real data challenges, such as noise, uncertainty, imbalanced classes, etc, 5) inefficiency in leveraging the representation of error, 6) inefficiency to scaling up well with large state-space or high-dimensional input space, and 7) inefficiency in handling latency, sparse rewards, and time dependencies properly.RL approaches suffer from the sixth and seventh mentioned issues. Abstraction approaches are proposed to mitigate these issues, but they bring other issues. For example, most of the abstraction methods require expert knowledge or cannot handle the complexities in multi-tasks and multi-agents domains. Classical electrocardiogram (ECG) classification approaches are time-series classification that suffer from the first, second, third, fourth, sixth, and seventh issues. Representation learning approaches have been used to mitigate the mentioned ones but they still cannot scale up well with the fourth and seventh issues properly and they require lots of training instances. Classification and clustering approaches for stationary datasets suffer from the first, second, third, fourth, fifth, and especially sixth issues. Representing learning approaches are proposed to handle the mentioned issues, but they require lots of training instances and do not capture all forms of available information. Current machine learning approaches in clustering, classification, and representation learning consider error as one scalar value while the error can bebreaking up into a vector of features and provide a high informative source of knowledge. In this dissertation, we provide novel approaches and perspectives for 1) abstraction for RL, 2) representation learning for ECG classification, and 3) classification and clustering. We present approaches in three separate parts in which they mitigate such issues.1) RL: RL approaches cannot be well scaled up to sparse reward and large state space domains i.e., the curse of dimensionality. Current efforts to mitigate sparse reward and curse of dimensionality issues are usually based on abstraction. Abstraction in a general form can be considered as a hierarchy that supports temporal and state abstraction such as decomposing the learning tasks into several sub-problems or a divide and conquer strategy. Hierarchical reinforcement learning (HRL) for task decomposition is a strong framework to use temporal abstraction, state abstraction, and subtask sharing in a structured way. The majority of current HRL techniques are based on assumptions of having experts with the knowledge of sub-goals to provide a correct hierarchy or having some in advance knowledge in other forms such as dynamic Bayesian networks. However, such assumptions can restrict the applications of these methods for autonomous learning where there is a limited expert's understanding. Extracting a hierarchical structure of learning tasks autonomously is the most known issue in HRL. The autonomous decomposition of tasks into a hierarchical structure holds the potential to significantly speed up learning, generalization, and transfer learning. In the second chapter of this dissertation, we address this problem for autonomous control and decision making in autonomous vehicle systems and propose a novel sequential association rule mining method to autonomously extract hierarchical structure of tasks in reinforcement learning in Markov decision processes (MDPs) and factored MDPs, called as SARM-HSTRL. The proposed method leverages association rule mining to discover the causal and temporal relationships among states in different trajectories, and extracts a task hierarchy that captures these relationships among sub-goals as termination conditions of different sub-tasks. We prove that. (Abstract shortened by ProQuest).

ISBN: 9798557010252Subjects--Topical Terms:

523869
Computer science.
Subjects--Index Terms:

Deep learning

Machine Learning-Based Decision Making in Autonomous Systems.
LDR:06729nmm a2200409 4500 001 2344610
005 20220531064607.5
008 241004s2020 ||||||||||||||||| ||eng d
020 $a 9798557010252
035 $a (MiAaPQ)AAI28256632
035 $a AAI28256632
040 $a MiAaPQ $c MiAaPQ
100 1 $a Ghazanfari, Behzad. $3 3683394
245 1 0 $a Machine Learning-Based Decision Making in Autonomous Systems.
260 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2020
300 $a 266 p.
500 $a Source: Dissertations Abstracts International, Volume: 82-06, Section: B.
500 $a Advisor: Fofanov, Viacheslav.
502 $a Thesis (Ph.D.)--Northern Arizona University, 2020.
506 $a This item is not available from ProQuest Dissertations & Theses.
506 $a This item must not be sold to any third party vendors.
520 $a Machine learning-based decision making is a statistical process of selecting different actions or assignments depending on input data. Autonomous decision making in machine learning (ML) is expected to handle different conditions such as non-linear environment, uncertainties in observations, and complexities of real-world domains in efficient and robust ways. Machine learning methods are categorized into reinforcement learning (RL), unsupervised learning approaches, and supervised learning methods. Clustering and classification are paradigms of unsupervised learning and supervised learning correspondingly. RL, classification, and clustering are applied depending on the type of problem. RL is considered as a learning approach for sequential decision-making problems. RL learns under uncertainty to maximize accumulated reward by trial and error, in which the reward signals may be received by delay. Classification and clustering are used for one-step decision-making problems. Classification and clustering assign instances to groups. There are labels in the classification problems while there are no labels in clustering problems. There are different types of static datasets such as images, documents, or classical ones, and non-stationary datasets like time-series.There are several issues for applying ML approaches in real domains: 1) expert knowledge is required to train approaches, which is not always cost-effective or feasible, 2) lack of generalizable approaches across multiple domains of ML, 3) inefficiency in handling complexities and non-linearity of patterns while there are considerable normal diversities in instances, 4) inefficiency in handling real data challenges, such as noise, uncertainty, imbalanced classes, etc, 5) inefficiency in leveraging the representation of error, 6) inefficiency to scaling up well with large state-space or high-dimensional input space, and 7) inefficiency in handling latency, sparse rewards, and time dependencies properly.RL approaches suffer from the sixth and seventh mentioned issues. Abstraction approaches are proposed to mitigate these issues, but they bring other issues. For example, most of the abstraction methods require expert knowledge or cannot handle the complexities in multi-tasks and multi-agents domains. Classical electrocardiogram (ECG) classification approaches are time-series classification that suffer from the first, second, third, fourth, sixth, and seventh issues. Representation learning approaches have been used to mitigate the mentioned ones but they still cannot scale up well with the fourth and seventh issues properly and they require lots of training instances. Classification and clustering approaches for stationary datasets suffer from the first, second, third, fourth, fifth, and especially sixth issues. Representing learning approaches are proposed to handle the mentioned issues, but they require lots of training instances and do not capture all forms of available information. Current machine learning approaches in clustering, classification, and representation learning consider error as one scalar value while the error can bebreaking up into a vector of features and provide a high informative source of knowledge. In this dissertation, we provide novel approaches and perspectives for 1) abstraction for RL, 2) representation learning for ECG classification, and 3) classification and clustering. We present approaches in three separate parts in which they mitigate such issues.1) RL: RL approaches cannot be well scaled up to sparse reward and large state space domains i.e., the curse of dimensionality. Current efforts to mitigate sparse reward and curse of dimensionality issues are usually based on abstraction. Abstraction in a general form can be considered as a hierarchy that supports temporal and state abstraction such as decomposing the learning tasks into several sub-problems or a divide and conquer strategy. Hierarchical reinforcement learning (HRL) for task decomposition is a strong framework to use temporal abstraction, state abstraction, and subtask sharing in a structured way. The majority of current HRL techniques are based on assumptions of having experts with the knowledge of sub-goals to provide a correct hierarchy or having some in advance knowledge in other forms such as dynamic Bayesian networks. However, such assumptions can restrict the applications of these methods for autonomous learning where there is a limited expert's understanding. Extracting a hierarchical structure of learning tasks autonomously is the most known issue in HRL. The autonomous decomposition of tasks into a hierarchical structure holds the potential to significantly speed up learning, generalization, and transfer learning. In the second chapter of this dissertation, we address this problem for autonomous control and decision making in autonomous vehicle systems and propose a novel sequential association rule mining method to autonomously extract hierarchical structure of tasks in reinforcement learning in Markov decision processes (MDPs) and factored MDPs, called as SARM-HSTRL. The proposed method leverages association rule mining to discover the causal and temporal relationships among states in different trajectories, and extracts a task hierarchy that captures these relationships among sub-goals as termination conditions of different sub-tasks. We prove that. (Abstract shortened by ProQuest).
590 $a School code: 0391.
650 4 $a Computer science. $3 523869
650 4 $a Remote sensing. $3 535394
650 4 $a Information technology. $3 532993
650 4 $a Artificial intelligence. $3 516317
650 4 $a Electrical engineering. $3 649834
653 $a Deep learning
653 $a Machine learning
653 $a Reinforcement learning
653 $a Representation learning
653 $a Error fission
690 $a 0984
690 $a 0544
690 $a 0489
690 $a 0800
690 $a 0799
710 2 $a Northern Arizona University. $b Electrical and Computer Engineering. $3 3560910
773 0 $t Dissertations Abstracts International $g 82-06B.
790 $a 0391
791 $a Ph.D.
792 $a 2020
793 $a English
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28256632