Boosting Artificial Intelligence in Software Defined Systems with - PowerPoint PPT Presentation

ý 2 Ñ � Ñ � ' f National University of Defense Technology Boosting Artificial Intelligence in Software Defined Systems with Open Infrastructure Li Zhou , Qi Tang, Jun Xiong, Haitao Zhao, Shengchun Huang, Jibo Wei College of Electronic Science, National University of Defense Technology, China Hunan Engineering Research Center of Software Radio, China Nov 15 th , 2018

Software Defined System ý 2 Ñ � Ñ � ' f National University of Defense Technology • 'software defined' is being labeled in a growing number of systems. software defined radio terminal software defined radar software defined mobile network software defined sensor network network software defined satellite network software defined cloud computing software defined data center cloud software defined storage software defined visualization 2

Software Defined System ý 2 Ñ � Ñ � ' f National University of Defense Technology • Software infrastructure plays an essential role. Type Commercial Open Source Open Standard Terminal Windows, iOS Linux, Android, ROS SCA, STRS Beacon, Floodlight, Network Cisco ACI, VMware NSX OpenFLow OpenDaylight, ONOS Microsoft Azure, Amazon Cloud Web Services (AWS), OpenStack, Kubernetes Alibaba Cloud … Components, Applications, Services Software Infrastructure Unified or standardized APIs Hardware 3

Software Defined System ý 2 Ñ � Ñ � ' f National University of Defense Technology Terminal SDS Cloud SDS Application Layer Service Layer Application 1 Application 2 Application N User Management Software Repository Application Develepment Comp Comp Comp Comp Comp Comp SDR Platform Controller SDR Platform Monitor Test and Certification ... Comp Comp Comp Comp Comp Comp Infrastructure Layer Comp Comp Comp Kubernetes Docker Ceph Registry Identity ... Infra API Computing Resource Pool Storage Resource Pool Network Resource Pool Infrastructure Layer Operating System Comp Comp Comp Comp Hardware Layer Hardware Layer Servers Switches Routers SDRs General Processors Hardware Devices CK P CK CK CPU RF Ethernet USB ... GPU FPGA ... P CK CK P P e.g. SCA e.g. Kubernetes • • Deployment, management, interconnection Automate deployment, scaling, and and intercommunication of software management of containerized services components 4

AI-able Terminal SDS ý 2 Ñ � Ñ � ' f National University of Defense Technology AI-able Terminal SDS Application Cognitive Engine • Repository Inspired by NASA Glenn cognitive communications systems project Cognitive Engine observe act observe act ... Orient Decide Training Knowledge Observe API Act API Sensing Waveform Transceiver Application Layer Application Application Application 1 Application 2 Application N Comp Comp Comp Comp Comp Comp ... Comp Comp Comp Comp Comp Comp Comp Comp Comp Wireless Environment Infra API Infrastructure Layer • Cognitive Engine: OODA Loop & Training Operating System Comp Comp Comp Comp • Possible Hardware: CPU, GPU, AI Chips HAL API • Hardware Layer Observe: General Processors Hardware Devices – query data from applications CPU USB ... GPU FPGA ... RF Ethernet spectrum data, parameters, etc . – • Act: – configure parameters into applications – Install/uninstall/start/stop/switch applications How to boost AI in SDSs with open infrastructure? 5

SDSDevOps Environment ý 2 Ñ � Ñ � ' f National University of Defense Technology Cloud Platform Intranet/ Internet Developments/Operations Terminal SDSs

Open AI-able Cloud Infrastructure ý 2 Ñ � Ñ � ' f National University of Defense Technology Open AI-able Cloud Infrastructure Basic Services • User Management Service Layer • Software Repository User Management Software Repository Application Develepment • Application Development SDR Platform Controller SDR Platform Monitor Test and Certification • SDR Platform Controller Learning Framework Data Analysis ... • SDR Platform Monitor Infrastructure Layer • Test and Certification Kubernetes Docker Ceph Registry Identity ... Computing Resource Pool Storage Resource Pool Network Resource Pool AI as a Service (AIaaS) • Learning Framework: Hardware Layer ‒ TensorFlow, Servers Switches Routers SDRs ‒ Caffe, CK P CK CK ‒ PyTorch P CK CK P P ‒ … • Data Analysis: ‒ Pandas, Cloud ‒ Statsmodels ‒ scikit-learn ‒ … 7

Open AI-able Cloud Infrastructure ý 2 Ñ � Ñ � ' f National University of Defense Technology Optimize cognitive engines in SDSs • Create training scenarios ‒ Centralized networks ‒ Decentralized networks ‒ Distributed networks ‒ Wired/wireless backhaul • Build dataset ‒ Geography data ‒ Spectrum data ‒ Channel data • Train parameters ‒ Neural network ‒ … • Train policies ‒ Support Vector Machine ‒ Reinforcement Learning ‒ … • Knowledge database ‒ Transfer learning 8

Deployment on NUDT Campus ý 2 Ñ � Ñ � ' f National University of Defense Technology Indoor SDRs Cloud Servers Outdoor SDRs • 60 SDRs are deployed initially ‒ 24 SDRs in the laboratory ‒ 36 SDRs outside in the campus • Cloud servers are deployed in the laboratory ‒ Intranet deployment only currently ‒ Internet deployment is on the way 9

Example Application (1) ý 2 Ñ � Ñ � ' f National University of Defense Technology e.g. Cell Deployment & Planning • Assumptions ‒ Spatial traffic (user distribution) varies from time to time in a region following some traffic patterns. ‒ Centralized controller & wired backhaul • Problems ‒ How to deploy minimal number of small cell base stations to meet users’ spatial traffic requirements? ‒ How to control the ON/OFF status of the small cell base stations to maximize energy efficiency? L. Zhou , Z. Sheng, L. Wei, X. Hu, H. Zhao, J. Wei & V. C. Leung, “Green Cell Planning and Deployment for Small Cell Networks in Smart Cities”. Ad Hoc Networks , vol. 43, pp. 30-42, June 2016. 10

Example Application (1) ý 2 Ñ � Ñ � ' f National University of Defense Technology e.g. Cell Deployment & Planning Cell Deployment & Planning Traffic Clustering • Our approach : Support Vector Machine (SVM) + Deep Neural Network 11

Example Application (2) ý 2 Ñ � Ñ � ' f National University of Defense Technology e.g. User Clustering • Assumptions ‒ Mobile ad hoc network R ‒ Users are divided into clusters CH2 distinguished by spectrum band (i.e., CH1 channels). ‒ Too large cluster size would result in heavy intra-cluster communication collisions ‒ Too small cluster size would result in large number of clusters, which leads to complex inter-cluster communication CH3 • Problem: How to group users into clusters with optimal size? • Our approach : Reinforcement Learning (POMDP) + Deep Neural Network 12

Summary and Future Work ý 2 Ñ � Ñ � ' f National University of Defense Technology • SDSDevOps is an environment that devotes to the development, operation, test and training of SDSs for students and researchers based on an open cloud infrastructure. • Massive centralized/Decentralized/distributed, wired/wireless backhaul scenarios can be created. • We plan to deploy and validate more applications that we used to study them by simulation on SDSDevOps. ‒ Cooperative Spectrum Sensing ‒ Cognitive MAC protocols ‒ Resource Management ‒ Mobility Management ‒ … 13