Encoder-Decoder LSTM

The core architectural principle involves two distinct LSTM networks. The encoder processes the input sequence of historical spectrum occupancy data (e.g., RSSI values over time) and compresses it into a fixed-length context vector. The decoder then takes this context vector and generates the output sequence, predicting future channel states step-by-step. This design explicitly handles variable-length input and output sequences, making it ideal for multi-step spectrum mobility prediction where the prediction horizon can be adjusted.

The fixed-length context vector, the final hidden state of the encoder, acts as a summary of the entire input history. This creates an information bottleneck, especially for long sequences. The decoder's performance depends entirely on this vector's ability to capture long-range temporal dependencies. To mitigate information loss, the attention mechanism was introduced as an enhancement, allowing the decoder to dynamically access all encoder hidden states rather than relying solely on the final context vector.

During training, the decoder is typically trained using teacher forcing. Instead of feeding the decoder's own previous prediction as the next input, the ground truth target value from the training data is used. This stabilizes training and accelerates convergence by preventing the model's errors from compounding over the prediction sequence. However, it creates a discrepancy between training and inference, a problem addressed by scheduled sampling, which gradually transitions from teacher forcing to using the model's own predictions.

This architecture excels at generating a prediction horizon of future channel states, not just a single next-step classification. The decoder can output a sequence of binary states (idle/busy) or probability distributions for each future time step. This capability is critical for proactive spectrum handoff, allowing a cognitive radio to evaluate the predicted spectrum availability window on a target channel and schedule a transmission burst before a primary user is forecasted to return.

Unlike standard feedforward networks, the LSTM cells within both the encoder and decoder contain gating mechanisms (input, forget, and output gates). These gates regulate the flow of information, allowing the network to learn dependencies over hundreds of time steps. This is essential for modeling primary user activity models that exhibit bursty traffic patterns, such as an MMPP, where the relevant historical context for predicting a future idle period may be far in the past.

Standard Encoder-Decoder LSTMs provide point predictions. To quantify prediction confidence, the architecture can be extended with Bayesian methods. By applying dropout as a Bayesian approximation during inference, the model can generate a predictive distribution over future channel states. This provides a confidence interval for the spectrum availability window, enabling a cognitive radio to make risk-aware handoff decisions, balancing the forced termination probability against unnecessary channel switching.

A recurrent neural network architecture using Long Short-Term Memory cells to capture long-range temporal dependencies in spectrum occupancy data. Unlike standard RNNs, the LSTM's gating mechanisms—input, forget, and output gates—mitigate the vanishing gradient problem, enabling the model to learn patterns across hundreds of time steps. This makes it ideal for modeling primary user activity that exhibits bursty, self-similar behavior over extended observation windows.

The specific future time step or lookahead window for which a spectrum mobility prediction engine forecasts channel occupancy. A short horizon (e.g., 1-5 time slots) reduces uncertainty but limits proactive handoff time, while a long horizon (e.g., 50+ slots) enables advanced reservation but accumulates prediction error. The encoder-decoder LSTM is specifically designed to handle multi-step prediction horizons by autoregressively generating future channel state sequences.

A stochastic framework used to mathematically represent the temporal behavior of licensed spectrum users. Common models include:

• ON/OFF exponential models: Memoryless arrival and departure processes
• Markov Modulated Poisson Process (MMPP): Captures bursty traffic with a hidden Markov state modulating the Poisson arrival rate
• Semi-Markov models: Allow non-exponential holding time distributions These models generate the training data for encoder-decoder LSTMs and define the ground-truth transition dynamics the network must learn.

A statistical model that infers unobservable channel occupancy states from observable signal measurements. In spectrum mobility, the true busy/idle state of a channel is hidden behind noisy sensor readings. The HMM uses Baum-Welch training and the Viterbi algorithm to estimate the most likely sequence of hidden states. While HMMs provide a Bayesian baseline, encoder-decoder LSTMs outperform them by learning non-linear temporal dependencies without requiring explicit state transition matrix specification.

A predicted temporal interval during which a specific frequency channel is forecasted to remain idle, enabling a secondary user to schedule a transmission burst. The encoder-decoder LSTM outputs a binary sequence indicating predicted idle (1) and busy (0) slots. Consecutive idle predictions define the availability window. Key metrics include:

• Window accuracy: How closely the predicted window matches the actual idle period
• Collision probability: The risk that a primary user arrives within the predicted window The window's length directly determines the maximum payload size a secondary user can transmit without interruption.

An online learning mechanism that detects and adjusts to statistical changes in primary user traffic patterns over time. Spectrum usage is non-stationary—traffic loads shift diurnally, new users appear, and interference patterns evolve. Without adaptation, a trained encoder-decoder LSTM suffers prediction degradation. Techniques include:

• Sliding window retraining: Periodically fine-tuning on recent observations
• Change point detection: Triggering retraining when the Hurst exponent or mean occupancy shifts significantly
• Elastic weight consolidation: Preserving previously learned patterns while adapting to new ones