Predicting consecutive spectrum opportunities with hidden Markov models

The increasing demands for wireless channels to accommodate the surge of internet of things devices and the associated services exacerbated the need for flexible channel allocation strategies. Opportunistic spectrum sharing is expected to provide a more reasonable use of the limited radio frequencies by allowing the coexistence of licensed users and unlicensed users in the same frequency. This arrangement is called opportunistic channel allocation, where unlicensed users explore the channel when the licensed user is not transmitting. The challenge in opportunistic spectrum allocation is to find transmission opportunities. Statistical and machine learning techniques have been used to forecast spectrum opportunities on time slotted channels based on acquired transmission data from licensed users. However, opportunity forecast is usually limited to one slot ahead of time. This work explores the use of Hidden Markov Model (HMM) training and predicting procedures to forecast a consecutive sequence of idle slots that can be explored for opportunistic transmission. To improve prediction accuracy and speedup the training process, the proposed training procedures reduce the classification alphabet, which is achieved by balancing the number of training sequences to limit the influence of outliers and provide opportunity forecast even when the training process is executed over a limited number of observed sequences. While an aggressive predictor may explore more opportunities as compared to a conservative predictor, the former may incur in higher collision rates while the latter may explore fewer opportunities. Hence, this work proposes a metric to assist in the process of selecting the best predictor's parameters to balance collision and opportunity rates. The proposed HMM predictors are evaluated on both synthetic and real traffic traces. The results on synthetic traffic shows that, for channel load from 30%$$ 30\% $$ up to 70%$$ 70\% $$, the proposed predictors identified over 80%$$ 80\% $$ of channel opportunities with the collision rates below 10%$$ 10\% $$. On real traffic traces, where the licensed users' transmissions have deterministic inter-packet delays, the HMM predictors identified over 95%$$ 95\% $$ of the channel opportunities with collision rates below 1%$$ 1\% $$.