Gaussian Process Regression

Unlike parametric models that assume a fixed functional form, GPR adapts its complexity to the data. It defines a distribution over functions directly, using a kernel to measure similarity between input points. This allows it to model complex, non-linear channel holding time distributions without pre-specifying a structure like an exponential or phase-type distribution.

The core advantage of GPR is its closed-form predictive distribution. For any future time point, it outputs both a predictive mean (the best estimate of idle time) and a predictive variance. This variance forms a prediction confidence interval, allowing a cognitive radio to quantify the risk of a collision and make risk-aware handoff decisions, directly optimizing the forced termination probability.

The behavior of a GPR model is entirely defined by its covariance function or kernel. This encodes prior assumptions about the signal's structure:

• RBF Kernel: Models smooth, infinitely differentiable functions.
• Matérn Kernel: A flexible class for modeling rougher, less smooth physical processes.
• Periodic Kernel: Captures repeating patterns in primary user activity. Composite kernels can be summed or multiplied to capture complex, structured primary user activity models.

Kernel parameters (e.g., length-scale, signal variance) and a noise variance term are learned from observed spectrum data by maximizing the log marginal likelihood. This Bayesian model selection principle automatically trades off model fit against complexity, providing a robust defense against overfitting that is built directly into the training objective, unlike methods requiring separate validation sets.

Standard GPR has a computational complexity of O(N³), making it prohibitive for large datasets. Sparse Gaussian Process methods, such as the Fully Independent Training Conditional (FITC) approximation, introduce a set of M inducing points (where M << N) to summarize the training data. This reduces complexity to O(NM²), enabling real-time spectrum availability window prediction on resource-constrained cognitive radios.

GPR provides a unifying probabilistic lens for spectrum prediction:

• It is a Bayesian interpretation of kernel ridge regression.
• A GPR with a specific kernel is equivalent to an infinite-width neural network.
• It generalizes the Kalman filter for state-space models in continuous time. This deep theoretical connection allows GPR to serve as a principled baseline for evaluating more complex LSTM spectrum predictors or Deep Q-Network handoff policies.

Gaussian Process Regression models the latent function governing spectrum occupancy over time, providing a predictive mean for the most likely channel state and a predictive variance that quantifies uncertainty. Unlike point-estimate models such as LSTMs, GPR outputs a full Gaussian distribution at each prediction point, allowing a cognitive radio to compute the probability that a channel will remain idle for a specified transmission window. This is achieved by placing a Gaussian Process prior over the function space and conditioning on observed spectrum sensing data using Bayes' rule.

• Kernel function selection: The choice of covariance function—such as a Matérn 3/2 or periodic kernel—encodes prior assumptions about the smoothness and periodicity of primary user traffic patterns.
• Hyperparameter learning: Kernel length-scales and signal variance are optimized by maximizing the log marginal likelihood of observed occupancy data, automatically adapting to traffic dynamics.

GPR's native uncertainty estimates enable risk-averse handoff policies that are impossible with deterministic predictors. A secondary user can set a confidence threshold—for example, requiring 95% posterior probability that a channel remains idle for the next 50 milliseconds—before initiating a transmission. If the predictive variance is high due to sparse recent observations, the radio can proactively switch to a target channel with tighter confidence bounds.

• Expected improvement acquisition: The radio evaluates candidate channels by computing the expected idle duration weighted by prediction confidence, selecting the channel that maximizes a utility function balancing throughput and collision risk.
• Adaptive sensing rate: When GPR uncertainty exceeds a predefined threshold, the radio increases its spectrum sensing frequency to gather more observations and reduce posterior variance before committing to a channel.

GPR extends naturally to multi-output regression through co-kriging and intrinsic coregionalization models, capturing correlations across both time and frequency dimensions. Adjacent channels often exhibit correlated occupancy patterns due to the spectral shape of primary user transmissions. By modeling the full spatio-temporal covariance structure, GPR leverages observations on one channel to improve predictions on neighboring channels.

• Coregionalization matrix: A positive semi-definite matrix encodes the covariance between channels, learned jointly with temporal kernel parameters during hyperparameter optimization.
• Transfer learning: A GPR model trained on a well-observed channel can inform predictions on a sparsely monitored channel through the shared latent function structure, reducing the cold-start problem in dynamic spectrum access.

GPR supports efficient sequential Bayesian inference, where the posterior distribution over the channel occupancy function is updated recursively as new spectrum sensing measurements arrive. This avoids retraining from scratch and enables real-time adaptation to non-stationary primary user traffic. The recursive formulation exploits the fact that the Gaussian Process posterior mean and covariance can be expressed as weighted combinations of kernel evaluations at observed points.

• Sparse Gaussian Process approximations: Techniques such as the Fully Independent Training Conditional (FITC) method use a set of inducing points to maintain a compact representation of the posterior, bounding computational complexity to O(M²N) where M is the number of inducing points and N is the number of observations.
• Forgetting factors: Exponential decay weights applied to older observations allow the GPR model to gracefully adapt to concept drift in primary user activity patterns without explicit change point detection.

GPR directly estimates the spectrum availability window—the contiguous future time interval during which a channel is predicted to remain idle with high confidence. By sampling from the GP posterior predictive distribution using Monte Carlo trajectory rollouts, the cognitive radio can compute the distribution of idle durations and select a transmission burst length that satisfies a target collision probability constraint.

• First-passage time analysis: The GP posterior is used to compute the probability distribution of the time until the latent occupancy function crosses a decision threshold, providing a principled estimate of the remaining idle period.
• Transmission scheduling: Given a payload of known size and modulation rate, the radio computes the required airtime and verifies that the predicted availability window exceeds this duration with probability above a configurable safety margin, such as 99%.

GPR can decompose complex spectrum occupancy patterns into interpretable additive components by constructing a composite kernel as a sum of specialized covariance functions. For example, a long-term smooth trend captured by a Radial Basis Function (RBF) kernel can be combined with a periodic kernel modeling daily usage cycles and a white noise kernel accounting for measurement error.

• Automatic relevance determination (ARD): Each input dimension—such as time-of-day, day-of-week, or adjacent channel occupancy—receives its own length-scale parameter, allowing the model to automatically learn which features are most predictive of spectrum availability.
• Changepoint kernel: A kernel defined as a product of a sigmoid transition function and a stationary kernel enables GPR to model abrupt shifts in traffic patterns while maintaining a fully Bayesian treatment of uncertainty around the changepoint location.