# 2 July 2020 - Daily Report

## Summary
### Expected Outcome :
* Understand how O-RAN works
* Understand why O-RAN use AI/ML
___
### Outcome :
* Understand the general concept of O-RAN
* Understand why and where O-RAN use AI/ML

### Further plan:
* Find the information of Acumos AI platform utilization in O-RAN

___
### Timeline
**08:55** - Writing Study Plan and Preparing to study
**09:00** - Starting to study
**12:58** - Praying and having a lunch
**13:33** - Continue study
**14:20** - Study the AI usage in RAN
**16:08** - Praying
**16:20** - Continue studying
**16.50** - Summarize works

---
## Study Notes

### Mobile Network Architecture
#### Typical Network
> Reference : https://www.brighttalk.com/webcast/16515/359818?utm_source=brighttalk-portal&utm_medium=web&utm_content=Parallel%20Wireless&utm_campaign=webcasts-search-results-feed


A Typical mobile network consists of Core Network where the data from different Access Network collected and can be exchanged each other. This network often called as a backbone. Typical mobile network also has an Access Network which give access to mobile devices and other devices so that it can access data or communicate. Each of the core network and access network consists of Hardware and Software part. Core network and access network are connected by transport network. On top of Core Network, there are different services that serve different purposes.

#### Disaggregated Network
In a disaggregated mobile network, hardware and software in the core and access network are separated. They now don't have dependencies on each other. O-RAN is a disaggregated network

### Disaggregation of Core Network

The main goals of core network disaggregation is to convert a dedicated hardware for a specific purposes into a software based function known as Network Functions Virtualisation. This enable us to run different function only using a COTS Server according to the reference. In this server, we use software to operate appropriately.
The simple example of this type of disaggregation can be found on modern phone. It is simple, small sized but has many different function that are embedded inside its software. We can access radio, music player, video player, etc in our phone. It can't be done two decades ago. We needed a dedicated hardware for each of that specific purpose back then.

### Disaggregation of Access Network


The picture above is a comparison between a traditional RAN. In the earlier RAN (left picture), the antennas are connected to the BBU (Base Band Unit) and RU (Radio Unit) by RF Cabling. This type of cabling can induce radio signal loss. Due to that, the later traditional RAN uses fiber cabling between the BBU to the RRU (Remote Radio Unit) which is located near the antennas. This approach will reduce RF cabling hence reduce signal loss. Aside from that, traditional RAN use proprietary hardware and software. It also has proprietary interfaces between BBU and RRU (in the later version).


Virtualized RAN (vRAN) is another type of RAN that uses COTS server and run proprietary software with virtualizer functions like the phone example before. This type of ran still uses proprietary RRU and proprietary interfaces between BBU and RRU.


O-RAN which is a disaggregated RAN uses open interface between BBU and RRU. Open interface means that O-RAN can accept any vendor software to operate in the interface hardware. It also uses GPP (General Purpose Processor) based COTS Hardware or so called SDR (Software Defined Radio) as the RRU. SDR is a radio device that its specific funtion is defined by the software inside it. SDR can be used in many purpose depending on the software installed. The BBU also uses COTS server like the vRAN.

### Why O-RAN?
#### Flexibility

Two pictures above show the software flexibility in O-RAN. Suppose that a mobile operator uses device A and software v1 is installed. The operator can easily install another software v2 in the same device A within the same network.

The picture above shows the hardware flexibility in O-RAN. Operator can easily upgrade/replace their hardware according to their needs. Each hardware could also be operated by different software.
With this level of flexibility, O-RAN has advantage over other type of RAN.

#### Cost Reduction
The picture above shows why service providers embracing O-RAN. Because of its flexibility, it avoids vendor lock-in. Many parts in the network that is easily replaceable so that any vendor can supply the parts. This leads to multiple vendors in a single networks. Any vendor that is participating in O-RAN need to make innovation in order to compete with other vendor. This offers different solutions depending on the requirement from service provider. Service in a rural area can be approached with different method from a dense populated area.
All of that features lead to cost reduction in network implementation and operation. Mobile operator can be easily scale their network and reduce time to market.
___

### AI and Machine Learning in Networks

> Reference : https://www.ericsson.com/en/reports-and-papers/ericsson-technology-review/articles/enhancing-ran-performance-with-ai

In the past few decades, computing power of a computer has increased rapidly. This allows people to develop better machine learning and AI model. It can be used to assist people works, automation, image recognition, etc. This great potential could also be applied in O-RAN.

There are 3 domains for RAN performance improvement that can be seen in the picture below

The first domain is network design. This domain focus on improving the parameter that define network deployment such as baseband, cells, RAN configurations and so on. This domain don't need a frequent update. It relies on network planning from the engineer such as when new cells are added to an existing network.

The second domain is network optimization. This domain focus on tuning the hyperparameter of the networks. The term hyperparameter in this case is different from hyperparameter term in machine learning. It refers to the parameter/input that control the behaviour of a network algorithm. The hyperparameter will affect output of the algorithm in a given situation. Some of the network entities affected by the hyperparameter tuning is cell clusters/indivitual cells. Its configuration such as static/semi-static is affected by the parameter

The RAN algorithms domain focuses on optimizing the L3 to L1 control parameter that directly affect the signal transmitted to/from the user. Because of the L3 to L1 algorithms adapt within a small time range, the parameter need to be adjusted in a fast timescale.

#### RAN optimization using AI

Controlling the network hyperparameter of RAN algorithms is safer than redesign all of the RAN algorithms. The network hyperparameter affect the output of RAN algorithms with a given input. This hyperparameter consists of outer control loop that does not modify the algorithms but only tunes its behaviour. In this case, AI is used in the second domain optimization which is network optimization.

#### The Non-RT RIC (RAN Intelligence Control)

>Reference:
>* [Ferlinda Feliana's Study Notes](https://hackmd.io/@ferlinda/SkUU8YXCL) about Non-RT RIC
>* O-RAN use cases and deployment, Feb 2020. It can be accessed [here](https://static1.squarespace.com/static/5ad774cce74940d7115044b0/t/5e95a0a306c6ab2d1cbca4d3/1586864301196/O-RAN+Use+Cases+and+Deployment+Scenarios+Whitepaper+February+2020.pdf)


One of the key principle of the O-RAN alliance is leading industry towards open, interoperable interfaces, RAN virtualization and big data and AI enabled RAN intelligence. O-RAN architecture shown in figure above contains Service Management and Orchestration (SMO) framework that also contain Non-RT RIC function. Its goal is supporting intelligent RAN ooptimization in non-real-time (i.e. greater than one second) by providing policy-based using data analytics and AI/ML training/inference. The non-RT RIC is one of the place where AI/ML model will be applicated to optimize the RAN. The other place is in the near-RT RIC parts.

According to Ferlinda's notes, there are other parts where we can deploy our model such as O-CU/O-DU. Following are the scenario for the AI/ML model deployment on O-RAN:
1. Deployment scenario 1.1: Non-RT RIC acts as both the ML training and inference host
2. Deployment scenario 1.2: Non-RT RIC acts as the ML training host and the Near-RT RIC as the ML inference host
3. Deployment scenario 1.3: Non-RT RIC acts as the ML training host and the O-CU/O-DU as the ML inference

Source : Ferlinda's study notes

Another interesting feature that i get from her notes is the training and deployment process of the AI/ML model. It can be depicted in the picture below.

The training and inference host can be located either in the same part or in a different part of the system. The learning process of the model ressemble the feedback in control system. Model training is done by using data which can be collected from output data of the system. After the training process is finished, the ML training host will update the model to the inference host.

## Conclusion
* O-RAN is a disaggregated network that its interface is open. This offers wide range of flexibility and gives a future-proof for new/existing network development
* O-RAN offers many benefits for the services provider
* AI/ML model is used in O-RAN to tune some network parameters in order to optimize network performance

###### tags: `Daily Report`