Artificial intelligence (AI) seems to be taking the world by storm. Every news media talks about it in many different contexts, from deep fakes to self-driving cars to ChatGPT. It is easy to read examples of business disruption driven by AI in almost every industry. It is somehow starting to look like “the only game in town,” the next big thing that is not “next,” but very current.
However, when we close our tablets or phones and get back to our day jobs, we do not normally see AI being deployed in every network at full scale, or taking care of automated network operations, or single-handedly managing CSP operations even in limited domains. The obvious question is: “If AI is already so good, why is not being used everywhere?”
Interesting questions do not normally have short, witty answers (that are good answers at least). However, we may start with another question: What does it mean to “use AI”?
Using AI at scale. Three preconditions
We see many AI solutions around in society today from image recognition to self-driving cars at full speed, which use different technologies such as Reinforcement Learning (RL) Natural Language Processing (NLP), or other AI technologies.
Nevertheless, the telecommunications sector faces the same three key challenges due to the rapid advancement of this technology and some preconditions need to be satisfied before the power of AI models is unleashed for full benefit.
- Trust: Trust in AI-driven insights and recommendations is crucial for their adoption and integration into operational processes. Transparency around the algorithms and data sources used to generate insights is key to building trust. Users should have visibility into how the AI models arrive at their recommendations, including the underlying logic and decision-making process. Progress toward closed-loop automation involves creating feedback mechanisms that allow the AI system to learn from the outcomes of its recommendations. This iterative process helps improve the accuracy and relevance of the insights over time, enhancing trust in the system's capabilities.
- Flexibility: Local requirements may vary across different geographical regions or markets. Therefore, AI-driven applications need to be flexible enough to accommodate these variations and tailor their recommendations accordingly. This could involve customizing models or adjusting parameters to align with specific needs. "Model drifting" refers to the phenomenon where the performance of an AI model degrades over time due to changes in the underlying data distribution or business environment. To address this, the system should be capable of detecting and adapting to such changes, including retraining a model trained initially on a dataset that incorporates a very diverse collection of network types from diverse markets, with local data to ensure its continued relevance and accuracy whenever is convenient to fulfil the business objectives.
- Agility: Integrating AI-driven applications into the communications service provider (CSP) operational framework requires agility to ensure seamless deployment and operation within existing infrastructure and processes. This may involve aligning with industry standards, interoperability requirements, and security protocols. Best CI/CD (Continuous Integration/Continuous Deployment) practices emphasize rapid, automated testing and deployment of software changes. Applying these principles to AI-driven applications enables quick iteration and updates, allowing for the seamless integration of new insights and recommendations into operational workflows.
With higher flexibility and agility of the AI-driven solutions, come easier and more efficient ways for service providers to deploy and operate AI capabilities. With more transparency of AI models comes a faster journey to have these models in production environments.
We know what we want to achieve. How do we do it?
How to boost Telco AI
We believe there is set of capabilities that will increase trust, flexibility and agility for service providers in their journey towards maximum value from AI. Explainable AI, tailored AI models and user interfaces, and a cloud native architecture, and our Ericsson Cognitive Software is focusing on fulfilling these preconditions in the specific domain of network design and optimization software. Let us explain what this means in practical terms.
Explainable AI to the rescue
Explainable AI is a set of methods and tools that build models to promote transparency and openness, allowing customers to understand the rationale behind algorithms and conclusions. Understanding and explaining AI models to users can help develop trust through increased transparency and awareness. Root Cause Reasoning (RCR) assists users in understanding the contribution of each key performance indicator (KPI) to the automated network evaluation and suggestions. RCR uses SHapley Additive exPlanations (SHAP) to explain the relative importance of KPIs within the cell. This in turn helps users to have complete visibility on the main contributing factors of network issues identified by the AI classifier. integrating explainable AI enhances decision-making within the telcos and beyond.
Delving deeper into the inner workings of AI will foster reliability, allowing massive AI adoption. Also, Explainable AI helps in mitigating bias and adjusting easily to AI regulations.
There is proven impact in applying e.g. RCR to network optimization. As per the report published in November 2023, Root Cause Reasoning was recently used by Swisscom and helped reach the highest ever recorded result globally in Umlaut Connect test.
Bridging the gap with tailored AI models and user interfaces
We see two sides to the flexibility challenge. AI model flexibility and application flexibility.
On model flexibility, Ericsson's network design and optimization solutions harness AI models trained on the largest and most diverse global datasets available in the industry. With them, customers benefit from extensive collective knowledge arising from these datasets. At the same time, there are some cases where these globally trained algorithms benefit considerably from taking into account local specificities like network topology. Automatic Cell Shaping is one of these cases. Retaining the flexibility for local adaptations when needed, service providers are empowered with complete control and flexibility to customize their AI model parameters, simplifying lifecycle management and reducing the time to market (TTM) of the latest capabilities.
On application flexibility, user interfaces and applications can be tailored to the needs of the specific operational procedures of the communication service provider, accommodating variations in workflows and user roles. These tailored workflows boost productivity and minimize time to action. With other functionalities like “bookmarks” for instance, users can create specific frequently used reports without configuring settings every time. All these contribute to reduce e.g. “number of clicks” and bring users faster to the desired outcome.
And this is a continuous improvement cycle. Feedback loops implemented in the system help us constantly improve user experience. Recent analysis indicates a considerable increase in user satisfaction of more than 20% with the introduction of these flexibility boosters.
Maximizing efficiency on a cloud-native architecture
Ericsson’s Cognitive Software architecture is cloud-native, allowing seamless integration with existing CI/CD pipelines, optimizing operational and infrastructure costs, and facilitating DevOps ways of working. The cloud-native architecture allows for In-Service Software Upgrades (ISSU) with zero downtime. Additionally, it enables rapid provisioning of trial and ad-hoc environments, reducing the timeframe for these projects from months to weeks.
Being modular, the Cognitive Software components can be scaled separately ensuring secure and updated systems. As we’ve covered just above, training and retraining of the models are essential for conducting smooth AI operations, and this would require scaling up and down periodically to accommodate the new data and computing power for model updates. Cloud-native facilitates these operations and sets itself as the most efficient architecture to host AI operations at scale.
Wrap-up: AI adoption can be accelerated, and we can help
By addressing the three major challenges for AI adoption, trust, flexibility and agility, we strongly believe service providers globally will reach new levels of efficiency and performance excellence. At Ericsson we intend to accompany service providers along the whole journey towards AI-driven autonomous networks, at any step of it with solutions and services for accelerating the value from the new capabilities.
Read more:
Transforming planning and optimization with AI – Ericsson
Network automation and AI – Ericsson
AI in Networks - Ericsson.
Explore Cloud Native.
