ORAN AIML flow Oct 2019 Costa Rica F

O-RAN - AI/ML flow Oct 2019 - Costa Rica F 2 F meeting Manoop Talasila – AT&T Guy Jacobson – AT&T

AI/ML in Amber Release O-RAN Model Lifecycle Acumos. AI Platform Agenda AI Integration Architecture ML Usecases Q&A © 2019 O-RAN Alliance All Rights Reserved

O-RAN AI/ML in Amber release RIC ML Features Why integrate RIC with Acumos? Key Objectives/Business Benefits • Quick models import into RIC through O 1 and adapt it into x. Apps (near-real time). • Deploy models as is into non-real time (mostly on SMO side). • Built adapter to speak RMR protocol. • Standardized platform, easy export, Docker deployment • RIC x. App developers can use Acumos AI catalog with ease • RIC x. App developers can focus on building great AI/ML based x. Apps rather than creating/maintaining infrastructure • Many companies have deployed/are using Acumos: Ericsson, Huawei, ZTE, Orange, AI 4 EU, AT&T, and Tech. M • Monetize and save on ML-based x. Apps by allowing many vendors to build and deploy x. Apps • Develop integration across Acumos, SMO (ONAP), and O-RAN © 2019 O-RAN Alliance All Rights Reserved 3

O-RAN - ML model lifecycle SMO ( ) ** ** Acumos is an LF project – upstream from O-RAN SC © 2019 O-RAN Alliance All Rights Reserved 4

What is Acumos? Create & On-Board Models Execute In Target Environment Runtime Systems AI Development Services and tools (Scikit-Learn, Tensor Flow, H 2 O, RCloud) Continuous Learning Model Local Learning Deploy Onboard AI Platform & Marketplace Train Executable Training Dataset Publish Search Predictor Data Sources Docker Images Review Chaining Rating Training / Testing Lifecycle Enhancing Model With Application Data Sets Sharing Models In Marketplace © 2019 O-RAN Alliance All Rights Reserved 5

Linux Foundation Acumos AI Members © 2019 O-RAN Alliance All Rights Reserved

RIC/Acumos Integration Architecture Model Catalog Deploy standard Acumos microservices Add new Acumos Microservice Adapter Two Docker images deployed together which speak GRPC/Protobuf, and have a self-contained request/response usage pattern Uses Python interface to standard RMR library to listen for and send messages on the bus to RIC using RMR protocol to serve as a single x. App via Helm chart Performance measurement Service Management & Orchestration ( docker registry ) O 1 RAN Intelligent Controller (RIC) near-RT x. App built using Acumos microservice x. App 1 x. App 2. . x. App# RMR Protocol suggest that the overhead involved is acceptable for non-realtime and many near-real -time x. Apps. x. APP / Acumos Microservice Adapter Python RMR Lib Config © 2019 O-RAN Alliance All Rights Reserved http /GRPC Standard Acumos ML model Microservice Kubernetes pod 7

Future Work Model Catalog Service Management & Orchestration ( AVOID OVERHEAD OF ADAPTER ML MODEL LANGUAGE SPECIFIC MODEL RUNNERS RMR MODEL RUNNER docker registry ) O 1 RAN Intelligent Controller (RIC) near-RT x. App built using Acumos microservice x. App 1 x. App 2. . x. App# RMR Protocol x. APP / Acumos Microservice Adapter Python RMR Lib Config © 2019 O-RAN Alliance All Rights Reserved http /GRPC RMR Protocol Standard Acumos ML model Microservice Kubernetes pod 8

RIC AI/ML Potential POCs Traffic Steering Qo. E Optimization MIMO Beamforming Optimization Training data from measurement reports Training data from Qo. E metrics Can create x. App to set policy here Reinforcement learning based on A 1/E 2 message? Needs many sources of training data: user location/context, RAN configuration/performance… x. App deployment through O 1 interface and configuration via A 1 interface APIs in O-RAN © 2019 O-RAN Alliance All Rights Reserved Control messages to optimize MIMO parameters 9

Questions? Useful Links RIC ML Based x. App - https: //wiki. o-ransc. org/display/RICA/ML-based+x. App Acumos - www. acumos. org Public Instance - market. acumos. org Acumos wiki - wiki. acumos. org Acumos Gerrit code repository - gerrit. acumos. org © 2019 O-RAN Alliance All Rights Reserved 10