Hands-On AI/ML, GenAI & Agentic AI in 5G Networks

Concepts

• 2G→5G evolution & architecture overview
• RAN: gNB, CU, DU, RU | Core NFs: AMF, SMF, UPF
• Network slicing concepts
• KPIs: RSRP, SINR, throughput, latency
• Where AI fits: SON, NOC, fraud detection

Hands-On Lab

• Explore 5G KPI CSV in Google Colab
• Plot RSRP and SINR distributions
• Identify poor cells using KPI thresholds
• Annotate an alarm log extract

Concepts

• Python: data types, loops, functions
• NumPy arrays & Pandas DataFrames
• Vectors, matrices, mean, std, variance
• Correlation — e.g. SINR vs throughput

Hands-On Lab

• Load KPI CSV; filter and group by cell
• Compute per-cell average SINR & throughput
• Plot SINR vs downlink throughput scatter
• Flag cells below KPI threshold

Concepts

• Supervised vs unsupervised learning
• Train / validation / test split
• Decision Trees & Random Forests
• Feature importance for network KPIs
• Use cases: congestion & outage prediction

Hands-On Lab

• Build classifier: 'Will this cell congest?'
• Feature engineering: rolling averages
• Train & visualise a Decision Tree
• Compare with Random Forest; check importance

Concepts

• K-Means clustering & elbow method
• PCA for dimensionality reduction
• Confusion matrix, precision, recall, F1
• ROC curve, AUC & cross-validation

Hands-On Lab

• K-Means on multi-cell KPI dataset
• PCA to visualise clusters in 2D
• Label clusters: 'overloaded', 'idle', 'edge'
• Evaluate S3 classifier with ROC-AUC

Concepts

• Perceptron → multi-layer ANN
• Loss functions & gradient descent
• RNN → LSTM gates: forget, input, output
• Time-series windowing for network KPIs

Hands-On Lab

• Build 2-layer ANN for congestion classification
• Create sliding-window sequences from KPI data
• Train LSTM for 15-min-ahead throughput prediction
• Plot predicted vs actual; compute RMSE

Concepts

• Attention mechanism: Query, Key, Value
• Multi-head attention & positional encoding
• Transformer encoder (BERT) vs decoder (GPT)
• Pre-training vs fine-tuning; tokenisation

Hands-On Lab

• Visualise attention weights on telecom log
• Use BERT tokeniser on alarm text corpus
• Run GPT-2 inference on alarm prompt
• Map Transformer components to their roles

Concepts

• Autoencoder: encoder, bottleneck, decoder
• Reconstruction error as anomaly score
• Threshold selection: percentile-based
• Network anomalies: spikes, faults, fraud

Hands-On Lab

• Train autoencoder on 6 weeks of normal KPI data
• Test on week with 3 injected anomalies
• Plot reconstruction error; find anomaly peaks
• Set threshold; compute precision and recall

Concepts

• ML approaches to network optimization
• SON: Self-Organizing Networks overview
• RL: state, action, reward, Q-value
• Q-learning algorithm & epsilon-greedy

Hands-On Lab

• Train regression model for optimal CIO values
• Simulate 3-cell RL environment (power control)
• Run Q-learning for 500 episodes
• Plot reward curve; compare epsilon values

Concepts

• Text preprocessing: stop words, stemming
• Bag of Words vs TF-IDF
• Named Entity Recognition (NER)
• Text classification for fault categories

Hands-On Lab

• Parse 500 syslog entries using regex
• TF-IDF vectorise alarms; train classifier
• Evaluate precision/recall per alarm category
• NER: extract cell IDs & error codes from logs

Concepts

• Scaling laws: why model size matters
• Pre-training & instruction tuning (RLHF)
• Emergent capabilities: reasoning, few-shot
• Context window, temperature, top-p, top-k

Hands-On Lab

• First API call: summarise an alarm report
• Experiment with temperature settings
• Zero-shot vs few-shot alarm classification
• Measure latency and token cost trade-offs

Concepts

• System, user, assistant prompt anatomy
• Zero-shot, one-shot, few-shot prompting
• Chain-of-thought (CoT) reasoning
• Structured output: JSON mode, XML tags

Hands-On Lab

• Write 3 versions of a fault diagnosis prompt
• Add CoT: 'Think step by step'
• Force JSON: extract alarm fields as structured data
• Prompt chain: extract → classify → recommend

Concepts

• Why LLMs hallucinate on domain facts
• RAG: embed, index, retrieve, augment, generate
• Embedding models & vector databases (ChromaDB)
• Evaluation: context recall & answer faithfulness

Hands-On Lab

• Embed a 50-page 5G vendor runbook
• Store embeddings in ChromaDB
• Query: 'How to resolve UPF heartbeat failure?'
• Compare RAG vs bare LLM on same question

Concepts

• Alarm flood reduction with GenAI
• Automated RCA report generation
• Shift-handover summary generation
• Config generation from natural language

Hands-On Lab

• Feed 30-alarm storm to LLM; extract root cause
• Generate shift-handover report from 8hr alarms
• Ask LLM to suggest CLI config for congested cell
• Build Streamlit UI for alarm summarisation

Concepts

• When to fine-tune vs prompt engineer vs RAG
• Supervised Fine-Tuning (SFT): instruction datasets
• LoRA & QLoRA parameter-efficient fine-tuning
• Evaluating: ROUGE, BERTScore, human

Hands-On Lab

• Format 100 alarm-to-RCA pairs as instruction data
• Run QLoRA fine-tuning on Mistral-7B
• Compare base model vs fine-tuned on alarms
• Calculate ROUGE score improvement

Concepts

• AI agent vs chatbot: what makes it agentic
• ReAct pattern: Thought → Action → Observation
• Tool calling / function calling in LLM APIs
• Memory types: in-context, vector, episodic

Hands-On Lab

• Define 3 tools: query_kpi_db, get_alarm_list, search_runbook
• Wire tools to LLM via function calling
• Run agent: 'Why is cell 42 showing high packet loss?'
• Observe ReAct trace; add memory store

Concepts

• Multi-tool agent for fault diagnosis
• Tool design: read-only vs action tools
• Guardrails & safety constraints for agents
• Human escalation triggers & auditability

Hands-On Lab

• Build fault diagnosis agent with 5 tools
• Test on 5 scenarios: UPF failure, interference, congestion…
• Guardrail: no config change without human approval
• Generate structured incident report per diagnosis

Concepts

• Why one agent is not enough
• Supervisor-worker pattern
• Shared state & message passing between agents
• Conflict resolution between agents

Hands-On Lab

• Build 3 specialist agents: Coverage, Capacity, Interference
• Wire to Supervisor Agent for mediation
• Simulate scenario where agents conflict
• Evaluate: multi-agent vs single general agent

Concepts

• ML model serving: Flask, FastAPI
• Edge AI in telecom: on-device RAN analytics
• LLM deployment: API, self-hosted Ollama/vLLM
• Monitoring: model drift, agent errors, hallucinations

Hands-On Lab

• Serve congestion classifier as FastAPI endpoint
• Pipeline: KPI stream → API call → alert
• Deploy local LLM with Ollama; call from agent
• Monitoring dashboard in Streamlit