OpenAI’s offerings—ChatGPT, Codex, Sora, and DALL-E—have caught the public’s imaginativeness and opened doors to galore opportunities for infusing AI into networks, systems, services, and applications. These cutting-edge AI technologies are present profoundly integrated with Microsoft products including Bing, Windows, Office, and Microsoft Teams. Within Azure for Operators, we are taking vantage of the important investments Microsoft has made and its expertise successful programming instauration models by processing method solutions that volition springiness our customers a competitory advantage. Our merchandise portfolio, which includes Azure Operator Nexus, Azure Operator Insights, and Azure backstage multi-access borderline compute is being augmented with generative AI technologies, empowering operators to efficiently lick real-world problems. But earlier we get into the solutions, let’s statesman with a little inheritance connected generative AI and caller AI advancements.

Background connected generative AI

OpenAI’s generative models person drawn important attraction for their exceptional show successful generating text, image, video, and code. Among these generative models, a notable breakthrough is generative pre-trained transformer (GPT), a ample connection exemplary with hundreds of billions of parameters. GPT is pre-trained connected a immense corpus of information from the unfastened internet, allowing it to comprehend earthy connection and make human-like responses to input prompts from users. ChatGPT, Codex (the exemplary down GitHub Copilot), Sora, and DALL-E are each derived from the pre-trained GPT (or instauration model). Codex is additionally trained connected codification from 54 cardinal GitHub repositories—a process known arsenic “fine-tuning.” To alteration the customization of GPT for caller connection tasks, OpenAI offers a paid API service that allows developers to fine-tune GPT connected domain-specific information done a command-line interface and query the fine-tuned exemplary without accessing the underlying model. Through a concern with OpenAI, Microsoft benefits from exclusive entree to the underlying exemplary and parameters of GPT, placing america successful a beardown presumption to make inference and a fine-tuning infrastructure. 

We person divided our AI and instrumentality learning investments into 4 categories:

1. Reactive management:​ Automated incidental management.​
2. Proactive management: Automated anomaly detection and responsibility localization.
3. AI and instrumentality learning infused into Azure for Operators products​.
4. AI and instrumentality learning engineering level crossed Azure for Operators​.

I privation to speech a small astir our investments that autumn nether the archetypal 2 categories. These systems showcase the imaginable of instauration models arsenic they are incorporated into our products, and they tin importantly interaction the mode mobile relation networks are developed, operated, and managed.

Reactive management: Data quality copilot for operators

Operators stitchery immense amounts of data, including node-level, gNodeB-level, user-level, and flow-level data, for purposes similar web monitoring, show tracking, capableness management, and debugging. In commercialized relation networks, the fig of specified counters and metrics that are regularly computed often exceeds respective thousands, accounting for tens of Gbps of information transfer. Retrieving applicable metrics and visualizing them is important for web operations. However, the complexity of modern wireless systems and the immense fig of counters progressive marque this task challenging, necessitating adept cognition to execute this indispensable operation.

The process contiguous involves specialists with adept cognition creating dashboards for a constricted fig of metrics, which the operators browse done to get applicable information. However, if operators necessitate customized data, specified arsenic visualizing throughput for a circumstantial idiosyncratic alternatively than aggregate throughput oregon if they request entree to a antithetic acceptable of metrics for analyzable debugging purposes, a loop done the specialists is required. The specialists request to place the applicable variables, constitute codification successful database query connection to harvester them successful an due manner, and past make and stock a dashboard. 

Can operators interact with their information by asking elemental questions successful earthy language, without having to retrieve immoderate of the analyzable antagonistic names oregon however to harvester them successful a database query language?

We judge that specified a strategy has the imaginable to importantly alteration the presumption quo. It would supply a much earthy mode to interact with relation information without dense reliance connected specialists. This would trim the clip to mitigate web issues, and it would supply much worth from the relation information by reducing the obstruction to customized insights.

The improvement of foundation models similar GPT-4 has importantly precocious the capabilities of earthy connection interfaces for information interaction, demonstrating singular show connected modular text-to-SQL datasets. Despite these achievements, challenges persist successful specialized and niche domains specified arsenic relation data. These challenges see the handling of specialized accusation that is often not publically available, the overwhelming measurement of information counters and metrics that exceeds the punctual size limits of these models, and the request for numerical accuracy that is important for decision-making successful web operations but which the instauration models are not adept at.

System architecture for information quality copilot for operators.

We person developed information quality copilot for operators, a earthy connection interface for retrieval and analytics tasks connected relation data, leveraging instauration models. It addresses the challenges posed by relation information done a operation of a domain-specific database with broad metrics descriptions, a semantic hunt for filtering applicable metrics wrong the models’ punctual size limits, few-shot learning for enhancing numerical accuracy successful codification generation, and adept feedback mechanisms that let for continuous betterment of the database done contributions from domain experts.¹ This copilot is being integrated into our Azure Operator Insights merchandise arsenic a cognition assistant.

Reactive management: Intent-based web absorption

Generally, relation networks are precise analyzable with absorption operations, heavy relying connected highly skilled professionals and blase absorption tools to create, update, and deploy web configurations. Configuration files tin beryllium respective tens of thousands of lines long. This process is not lone labor-intensive but besides error-prone, underscoring a request for automation to alleviate the absorption load for web operators.

We person been moving connected a promising paradigm called intent-based networking (IBN), a solution to simplify web absorption for operators. It allows web operators to specify the desired behaviour oregon “intent” of their web successful earthy language. They tin say, “Allow ICMP postulation successful my network,” and past the solution automatically translates the intent into updated web configurations. IBN tin contiguous these updated configurations to web administrators for reappraisal anterior to their deployment, ensuring web information portion keeping minimal quality intervention.

Intent-based networking cause powered by GPT simplifies web management.

Although the conception of IBN has been astir for immoderate time, its implementation has been hindered by the complexities of earthy connection knowing and the intricate task of configuration generation. Motivated by caller advances successful generative AI (for illustration GPT), we revisited this occupation and developed a instrumentality named “IBN agent” based connected GPT. Our IBN cause takes arsenic input the moving web configuration and the user’s earthy connection intent. It past queries GPT to update the web configuration according to the idiosyncratic intent. Utilizing existing configuration syntax checks and web verification tools, the IBN cause besides flags errors successful the GPT-generated configurations. Moreover, users tin intervene astatine immoderate constituent and supply feedback connected immoderate undesired behavior. Based connected these identified errors oregon idiosyncratic feedback, the IBN cause iteratively refines the configuration with GPT until each automated and quality checks are passed. We judge that IBN holds important imaginable to simplify web configuration successful the future.

Proactive maintenance: Next procreation communications copilot for operators

Practitioners, engineers, researchers, and students tin find themselves grappling with a multitude of acronyms and intricate terminology with accusation dispersed crossed galore documents, which makes working with and processing standards-compliant systems an onerous and time-consuming task. For example, an engineering squad moving connected implementing a registration petition process arsenic a portion of gathering 5G virtual halfway would request to place each the applicable method specifications from among thousands of documents and recognize the telephone travel and connection formats arsenic described successful those specifications.

The existent method of acquiring this accusation involves sifting done galore webpages and method specification documents. While this attack provides extended comprehension of a taxable from assorted sources, it tin besides beryllium time-intensive and tedious to identify, gather, and synthesize accusation from aggregate applicable sources.

Foundation models correspond a important advancement successful providing synthesized, readily comprehensible answers to idiosyncratic queries related to wireless connection specifications. However, contempt the usefulness of state-of-the-art ample connection models, they besides nutrient irrelevant oregon inaccurate responses to galore queries related to niche and specialized domains.

We person developed a conversational AI instrumentality for accusation synthesis of wireless connection specifications.

Like ChatGPT, the nextgen communications (NGC) copilot offers a question-and-answer interface, but with an enhanced quality to supply much close and applicable answers connected topics pertaining to wireless connection method specifications. NGC copilot builds connected instauration models, punctual engineering, and retrieval augmented procreation approaches; it features a domain-specific database, tailored connection embeddings, and a idiosyncratic feedback mechanism. For much close responses, it integrates into its database method specifications and standards that are often overlooked by accepted models owed to their niche nature. The strategy uses a specialized word-embedding exemplary to amended recognize telecom jargon, improving its query effect relevance. Experts tin besides supply feedback, which helps refine the database and amended reply quality. We person been piloting NGC wrong our engineering teams and its show has been excellent.

Proactive management: Network configuration anomaly detection

One of the astir communal causes of web disruptions contiguous is web configuration errors. Configuration governs the protocols and policies that modulate and power web access, performance, security, billing, and more. Misconfigurations, erstwhile they occur, tin pb to a frustrating idiosyncratic acquisition with dilatory performance, deficiency of connectivity, oregon adjacent sweeping work outages. Operators who acquisition specified outages often endure from nonaccomplishment of estimation and revenue.

Despite the value of close web configuration, configuration absorption contiguous remains a situation for operators. Manual adjacent reappraisal of configuration changes tin person constricted effectiveness. Device configurations are often low-level, complex, and long—making them notoriously challenging to audit manually and astatine scale. On the different hand, automation is besides not a panacea; it’s prone to errors, bugs, and mistakes.

The configuration anomaly detection investigation pipeline.

Many configuration errors are evident successful hindsight and could beryllium detected by sufficiently intelligent learning models. For this reason, we person invested successful processing AI-driven anomaly-detection tools that tin proactively place and artifact erroneous configuration changes earlier they are applied to the network—before they tin interaction existent users. Machine learning is adept astatine identifying communal configuration usage patterns and anti-patterns. It tin efficaciously sift done changes to disregard those that are intentional and alert operators astir those that are apt unintentional oregon erroneous.

Given a postulation of akin configuration files (such arsenic JSON, XML, oregon YAML), our strategy synthesizes a communal template that captures the similarities betwixt these configurations, leaving placeholders for differing values. Using the synthesized template, our strategy employs a state-of-the-art, unsupervised anomaly-detection technique, known arsenic the isolation forest, to pinpoint apt errors successful configurations. These imaginable anomalies are reported with an anomaly-likelihood people for review. In this way, we purpose to assistance operators with harmless and reliable absorption of their 5G networks by leveraging automated validation of configurations. For real-world scenarios and further method details, delight work our caller paper.²

Microsoft liable AI

We recognize that AI and instrumentality learning-based solutions whitethorn impact ethical concerns regarding the underlying models, their grooming data, and associated biases. To code these concerns, the office of liable AI shepherds the AI projects astatine Microsoft connected hazard appraisal and mitigation. We enactment hard to recognize the aspects that necessitate betterment regarding bias and discrimination, and we strive to person wide support connected compliance. We walk connected each guidelines to the engineers to guarantee liable usage without slowing progress.

Explore the Microsoft portfolio of products

My anticipation is that these examples amusement that instauration models importantly heighten the Azure for Operators portfolio of products. There is simply a batch much to say, and determination are galore further examples of systems we person developed, but I volition permission that for different time.

---

¹Microsoft, Adapting Foundation Models for Operator Data Analytics, Manikata Kotaru, HotNets 2023.

²Microsoft, Diffy: Data-Driven Bug Finding for Configurations, Siva Kesava Reddy Kakarla, Francis Y. Yan, and Ryan Beckett, April 2024.

The station Empowering operators done generative AI technologies with Azure for Operators appeared archetypal connected Microsoft Azure Blog.