Automating AI solutions for Telecom: A blueprint for industry-wide innovation
By Manisha Kalita
RIC Architect Rakuten Symphony
August 6, 2024
In today’s world, AI solutions are everywhere, whether as a marketing tactic or a technological breakthrough. Regardless of the context, end users often remain unaware of the complexities behind AI, such as the specific machine learning models used, the massive data sets processed for training, and the significant computational power needed to run them.
For example, a healthcare provider is interested in using an AI system to assist in analyzing medical images. They are usually unaware with the specifics of how convolutional neural networks or deep learning algorithms work. Their focus is on the outcome: an AI system that can accurately identify potential health issues in medical images, aiding doctors in making quicker and more accurate diagnoses.
This hands-off approach of the end users is logical and justified given the high demands for quick and efficient solutions. With this division of responsibilities, the creation and fine-tuning of AI models are taken care by experts who can develop models that are accurate, efficient and reliable. Meanwhile, end users can focus on leveraging the benefits of AI in their respective domains.
The unseen gap: Automatic model selection
At present, selecting the best AI model for a specific problem often involves a manual process. Experts evaluate different models based on performance metrics, computational efficiency and suitability for the specific task at hand. This selection process can be time-consuming and requires a deep understanding of machine learning and the specific application domain.
With the growing demands for AI solutions across various use cases and the abundance of AI model providers, it becomes increasingly tedious to manually identify the most suitable model for each use case. Soon, there will be a need for systems that can automatically select the most appropriate AI model(s) for a given problem. Such systems would simplify the deployment of AI solutions, making them more accessible to non-experts. Automated model selection could involve evaluating models based on predefined criteria, using meta-learning techniques, or leveraging AI to recommend models based on historical performance data.
AI usage in telecom with Open RAN
In today’s telecom sector, Open RAN (Radio Access Network) is no longer just a buzzword; it is now a significant part of the industry's landscape. According to recent research by Dell’Oro Group, Open-RAN is projected to account for 7% to 10% of global RAN revenues in 2024, with expectations to grow to 20% to 30% by 2028. Telecom operators are willing to try out multi-vendor integration to enhance flexibility, innovation and to leverage a diverse range of solutions from different vendors and thus promoting collaborative solution and driving technological advancements.
AI for telecom network optimization
AI plays a crucial role in solving problems and optimizing resources in telecom networks. AI models can be used for tasks such as traffic prediction, anomaly detection, resource allocation and network performance optimization. Due to the rapidly changing network scenarios, relying on human efforts to find issues and identify the appropriate model can be time-consuming and inefficient. To fully realize the potential of AI in a multi-vendor environment, there is a need for a standardized interface that allows operators to programmatically identify and select AI models from a catalogue. At the same time, this catalogue would include models developed by various vendors, each designed to address specific network challenges.
The Benefits of a Standardized Interface
Interoperability: A standard interface ensures that AI models from different vendors can seamlessly integrate into the telecom network, regardless of the underlying infrastructure.
Ease of Integration: Operators can easily incorporate new models into their network operations without extensive customization or configuration.
Scalability: The ability to programmatically select and deploy models allows for scalable and efficient network management, adapting to changing demands and conditions.
Vendor Collaboration: A standardized model registration process encourages collaboration among vendors, leading to a richer and more diverse set of AI solutions.
Optimizing RAN with AI/ML: Model Registration, Selection, and Deployment in Open-RAN
RAN Intelligent Controller (RIC), evolved with Open RAN, is a cloud-native component designed to control and optimize RAN functions. Aligned with the core principles of Open-RAN, RIC enables multi-vendor interoperability, intelligence, agility, and programmability in radio access networks through third-party applications. These applications often use AI/ML-based recommendations or predictions to make decisions that optimize network performance.
To provide interoperability, O-RAN ALLIANCE working groups are focusing on standardizing the interface to use the AI/ML models. The usage of AI/ML models involve the following steps:
Model Registration: Vendors can register their AI models in a centralized catalogue, providing details about the model’s capabilities, performance metrics and compatibility requirements.
Programmatic Model Selection: Operators can use APIs to programmatically search and select models from the catalogue based on specific criteria such as problem type, performance benchmarks and resource requirements.
Integration and Deployment: Once selected, the models can be integrated into the network management systems using standardized interfaces, ensuring smooth deployment and operation.
Continuous Improvement: Feedback loops can be established to continuously monitor the performance of deployed models, allowing vendors to refine and improve their solutions based on real-world data.
Efficient AI Model management with standard interfaces
An application can leverage the standard interface to discover, deploy and use AI models efficiently. Here’s how it works:
Discovering Models: The application uses the standard interface to search the centralized catalogue for suitable AI models. It can query models based on specific criteria such as problem type, performance metrics and resource requirements.
Deploying Models: Once the appropriate models are identified, the application calculates the required resources based on the desired performance, such as the number of inference requests per time period. The standard interface facilitates the deployment of these models by integrating them into the network management systems seamlessly.
Utilizing Models: The application can use deployed models to achieve its goals. It is not tied to any specific model and can be developed by different vendors. The application identifies and selects the most appropriate models using a standard interface and may use multiple models simultaneously for better accuracy and reliability. For instance, to predict network resources, it might combine outputs from traffic prediction and anomaly detection models, aggregating these predictions to optimize resource allocation.
https://symphony.rakuten.com/blog/...int-for-industry-wide-innovation
|
Thread abonnieren
