Online Learning Model

Unlike batch learning, online models update their weights one sample at a time or in mini-batches as new telemetry arrives. Each update computes a gradient based solely on the incoming data point, adjusting the model's internal state without revisiting historical data. This enables adaptation to concept drift—the phenomenon where the statistical properties of network traffic change over time due to events like stadium crowds or new cell site activations. Common algorithms include Stochastic Gradient Descent (SGD) with a learning rate that controls update magnitude.

Online models consume data from unbounded streams rather than static datasets. In a RAN context, this stream consists of real-time telemetry such as PRB utilization, CQI reports, and RRC connection counts flowing from E2 nodes. The model processes each observation exactly once—a property called one-pass learning—making it computationally efficient for high-velocity data. This contrasts with batch training, which requires multiple epochs over a stored dataset and cannot adapt between full retraining cycles.

A defining capability of online learners is their inherent resilience to concept drift—changes in the relationship between input features and target variables. In cellular networks, drift manifests as:

• Sudden drift: A flash mob creating an unexpected traffic hotspot
• Incremental drift: Gradual subscriber growth in a residential area
• Recurring drift: Daily commute patterns shifting seasonally Online models continuously track these shifts, unlike static models that grow stale and require manual retraining. Techniques like exponential forgetting assign higher weight to recent observations.

Online learning eliminates the need to store massive historical datasets or provision large GPU clusters for periodic retraining. Each update is O(1) in memory relative to dataset size—the model maintains only its current parameters and processes one sample at a time. This makes online models ideal for deployment directly on Near-RT RIC platforms or edge nodes with constrained resources. The computational cost per update is constant, enabling predictable latency budgets suitable for control loops operating on 10ms to 1s timescales.

Online models operate in a continuous predict-then-update cycle:

1. Predict: The model forecasts a target (e.g., cell load in 5 minutes) using current parameters
2. Act: The RIC executes a load balancing decision based on the prediction
3. Observe: The actual outcome arrives as new telemetry
4. Update: The model computes the prediction error and adjusts parameters to reduce future errors This closed-loop architecture mirrors the Observe-Orient-Decide-Act (OODA) paradigm and is foundational to Self-Organizing Networks.

Online learning is theoretically grounded in regret minimization—the goal of performing nearly as well as the best fixed model in hindsight. Regret measures the cumulative difference between the online model's predictions and those of an optimal static model. Algorithms like Online Gradient Descent and Follow-the-Regularized-Leader (FTRL) provide formal guarantees that regret grows sublinearly over time, meaning the average performance converges to optimal. This mathematical foundation assures operators that the model will not catastrophically diverge.

A phenomenon where the statistical properties of the target variable change over time, violating the i.i.d. assumption of static models. In RAN environments, concept drift occurs when user mobility patterns shift due to new urban infrastructure, seasonal events, or network reconfiguration.

• Sudden drift: An abrupt change, such as a stadium event ending
• Incremental drift: Gradual shifts like suburban population growth
• Recurring drift: Cyclical patterns like weekly commuter behavior

Online learning models are inherently designed to adapt to concept drift without manual intervention, making them superior to batch-trained models for dynamic cellular environments.

The foundational optimization algorithm that enables online learning by updating model parameters one sample at a time. Unlike full-batch gradient descent, SGD computes the gradient using a single incoming telemetry record, allowing the model to adapt continuously.

• Learning rate controls the step size and adaptation speed
• Momentum smooths updates by accumulating a velocity vector
• Mini-batch SGD processes small groups of samples for stability

In predictive load balancing, SGD-based online learners ingest streaming PRB utilization metrics and CQI reports to incrementally refine cell load forecasts without storing historical data.

The automated monitoring process that identifies when a deployed model's predictive performance degrades beyond acceptable thresholds. While online learning models self-adapt, drift detection serves as a safety mechanism to trigger alerts or fallback strategies.

• Population Stability Index (PSI) measures feature distribution shifts
• Prediction error monitoring tracks rolling RMSE or MAE
• Data quality checks detect missing or corrupted telemetry streams

When drift is detected, the system may increase the learning rate, reset the lookback window, or temporarily revert to a rule-based load balancing policy.

The length of time into the future for which the online learning model generates a forecast. This parameter directly impacts the proactivity of load balancing actions and the model's accuracy.

• Short horizon (1-10 seconds): High accuracy, enables reactive handover tuning
• Medium horizon (10-60 seconds): Balances accuracy with proactive resource allocation
• Long horizon (1-5 minutes): Enables network-wide traffic steering but with higher uncertainty

Online models dynamically adjust their prediction horizon based on observed temporal volatility in cell load patterns, using shorter horizons during unstable periods.

The fixed length of historical time-series data used as input for each incremental prediction. In online learning, the lookback window defines the temporal context the model can observe before updating its parameters.

• Sliding window: Continuously shifts forward as new data arrives
• Exponentially weighted: Recent samples receive higher importance
• Adaptive sizing: Window length adjusts based on detected seasonality

A well-tuned lookback window captures diurnal traffic patterns while remaining short enough to allow rapid adaptation when concept drift occurs.

A core Federated Learning algorithm that enables online learning across distributed base stations without centralizing raw telemetry data. Each base station trains a local model incrementally, and only encrypted parameter updates are shared.

• Local SGD steps are performed on each base station's streaming data
• Periodic aggregation averages model weights on a central server
• Differential privacy noise can be added to protect user patterns

FedAvg allows online learning models to benefit from network-wide patterns while maintaining data sovereignty and complying with privacy regulations.