GenAI for telecom OSS

This post explores the use of GenAI in transforming service life cycles and service operations in Telecom OSS.

The use of AI starts with identifying and prioritising use cases to be defined that will result in the implementation of transformed service lifecycles and service operations that use AI or GenAI.

Identifying and prioritising use cases for OSS

One way of identifying use cases for ML, AI, and GenAI in Telecoms is by considering the sources of data available from CSPs.

These sources of data are available at the following layers

Data Collection and Data Flow Layers

This includes

  • Data collected from Network Elements, Network Element Management Systems per Technology Domain
  • Data collected from Customer Devices, and Physical and Virtual Customer Network elements
  • Data entering and leaving the network via Carrier Interconnect
  • Data traversing internally within the CSP Network (Access, Aggregation and Core)
  • Data leaving customer sites, traversing the access, aggregation and core networks, entering and leaving data centres and cloud.
Data Aggregation, Processing, Control/Routing Layers

This include

  • Data Aggregated across Technology Domains – Network Device and Connectivity data from related platforms and Cloud platform data
  • Control Data from Network Controllers that work across domains with respective Domain Controllers
  • Transactional Data from Network Automation used for network configuration, network provisioning, and service activation.
Service Orchestration, Service Assurance Layers

This includes

  • Monitoring and performance data from Network Services and Customer Facing Services
  • Service availability and uptime metrics
  • Performance metrics of automated processes
  • Success and failure rates of automated tasks
Selection and use of Pre-Trained models

General purpose LLMs are in most cases not suitable to drive improvements in the use cases identified using the above.
The inferences provided by such models mostly do not benefit the cost of using them. Telecom and OSS tasks are specific to multiple technology domains like DWDM, Carrier Ethernet, Wireless, and IP.
Hence pre-trained models on telecom data provide a head start toward the use of such models in the use cases described above, in the context of relevant business scenarios.

Identifying and selecting pre-trained models requires considerations of cost to obtain a license and cost to run, costs to customise and costs to fine-tune. Additional considerations include the type of model to use for the type of problem. e.g. Transformer type models used in LLMs or “simpler” models that use Decision Trees, Linear/Logistic Regression for supervised learning, K-NN for unsupervised learning or Feedforward Neural networks.

Additionally, as LLMs continue to evolve, the commercially available model could be updated by the vendor. So inferences made previously might change significantly and these outputs may be significantly different requiring altering of prompts previously used.

Fine Tuning Models for Telecommunications use

Generic models can be fine-tuned for telecom by training on telecom-specific Q&A types of texts or prompts. A final stage would use reinforcement learning from human feedback.

Other Challenges and Considerations

One of the challenges with using Gen AI is remaining non-biased and this requires vigilance when using data sets to avoid the use of skewed data sets.
Another challenge is understanding the training algorithms and training parameters, which means understanding how the models are trained to help prevent bias creep and misinterpreted results.
A third challenge is handling data privacy, especially with the use of customer data.
And a fourth challenge is keeping up with regulatory laws as they are updated.
Future posts to elaborate on the above.