Resource Scheduling and Performance Optimization in Distributed Environment Based on Federated Learning and Reinforcement Learning
- Alternative Title
- 연합학습 및 강화학습 기반 분산 환경에서 자원 스케줄링 및 성능 최적화
- Abstract
- 본 논문에서는 딥러닝 기반 분산 통신 및 레이더 환경에서 자원 스케줄링과 성능 최적화 기법을 제안합니다. 먼저 연합학습(Federated learning) 기반 시스템에서 분산 환경을 고려합니다. 다중 클라이언트가 서버에 gradient를 전송할 때 학습 효율을 높이기 위한 전략을 제안합니다. 첫 번째로, 각 클라이언트의 송신 신호가 겪는 채널 페이딩으로 인한 왜곡을 극복하기 위해 채널 이득을 이용한 MRC (Maximum Ratio Combining) 기반의 gradient update 가중치 설정, 채널이 좋지 않은 시점의 학습을 차단하는 임계값 설정 전략과 각 클라이언트의 송신 전력이 제한된 경우 중요한 정보를 담은 송신 데이터(크기가 큰 gradient 벡터)에 적응적으로 전력을 할당하는 기법을 제안합니다. 두 번째로, 각 클라이언트의 송신 신호가 겪는 채널 페이딩으로 인한 왜곡을 극복하기 위해 딥러닝의 한 종류인 Auto-Encoder 기반 송, 수신 전략을 제안합니다. 제안된 학습 및 송, 수신 전략은 채널 보정 및 성능 향상에 효과적임을 확인하였습니다. 그 후 OFDM 기반 RadCom 시스템에서 non-linear stepped OFDM 기술과 신호처리 기법을 제시한 뒤 강화학습 기반 분산 부대역 할당 기법을 제안한 뒤 그 성능을 검증합니다.
In this paper, the resource scheduling and performance optimization methods in deep learning-based distributed communication and radar environments are proposed. First, a distributed environment in a federated learning-based system is considered. When multiple clients send gradients to the server, we propose a strategy to increase learning efficiency. First, to overcome the distortion caused by channel fading experienced by each client's transmission signal, we set the gradient update weight based on MRC (Maximum Ratio Combining) with the channel gain and set the threshold value to block corresponding round at the time when the channel is not good. We propose a strategy and a technique for adaptively allocating power to transmit data containing important information, that is, large gradient vector when the transmit power of each client is limited. It has been confirmed that the proposed training and transmitting/receiving strategies are effective for channel correction and performance improvement. After that, the non-linear stepped OFDM signal and processing method are presented in the OFDM-based RadCom system, and the reinforcement learning-based distributed subband allocation method is proposed and the performance is verified.
- Author(s)
- Yunji Yang
- Issued Date
- 2022
- Awarded Date
- 2022. 2
- Type
- Dissertation
- Publisher
- Pukyong national university
- URI
- https://repository.pknu.ac.kr:8443/handle/2021.oak/24425
http://pknu.dcollection.net/common/orgView/200000602248
- Alternative Author(s)
- 양윤지
- Affiliation
- Pukyong National university, Graduate school
- Department
- 대학원 스마트로봇융합응용공학과
- Advisor
- Jaehyun Park
- Table Of Contents
- I. Introduction 1
II. Efficient Gradient Updating Strategies for Federated Learning OverWireless Backhaul 3
2.1 Introduction 3
2.2 System model for federated learning over wireless backhaul 6
2.2.1 CNN architecture for image recognition 7
2.2.2 Signal model for wireless backhaul 9
2.2.3 Federated Learning for image recognition 10
2.2.4 Zero-forcing linear gradient estimation 12
2.3 Federated learning for motion recognition using MD signatures 14
2.3.1 System model for distributed MD radars with wireless backhaul 14
2.3.2 Deep learning aided gradient estimation over wireless backhaul 18
2.4 Gradient updating and adaptive power allocation strategies for the federated learning over wireless backhaul 21
2.4.1 Gradient update by maximum ratio combining 21
2.4.2 Binary gradient update by thresholding 22
2.4.3 Adaptive power allocation strategy based on the gradient information 22
2.5 Experimental Results 24
2.5.1 Experimental results of deep learning aided gradient estimation in distributed MD radar environment 27
2.5.2 Experimental results of gradient updating and adaptive power allocation strategies in a distributed client environment 30
2.6 Conclusions 38
III.Stepped-carrier OFDM Joint Radar and Communication System with Nonlinear Hopping Pattern 39
3.1 Introduction 39
3.2 Joint radar and communication system model for stepped-carrier OFDM 41
3.2.1 Tx signal model 42
3.2.2 Rx signal model for radar receiver 42
3.2.3 Rx signal model at information receiver 43
3.3 Radar processing with nonlinear hopping pattern 45
3.3.1 Nonlinear stepped OFDM radar processing scheme 45
3.3.2 Proposed Radar processing scheme with calibration 48
3.4 Subband selection for stepped-carrier OFDM with nonlinear hopping pattern 49
3.4.1 Centralized scheduling 49
3.4.2 Decentralized scheduling 50
3.5 Experimental Results 53
3.6 Conclusions 59
IV. Conclusions 60
Acknowlegement 66
- Degree
- Master
-
Appears in Collections:
- 대학원 > 스마트로봇융합응용공학과
- Authorize & License
-
- Files in This Item:
-
Items in Repository are protected by copyright, with all rights reserved, unless otherwise indicated.