Neural Network Predictor for Fast Channel Change on DVB Set-Top-Boxes.

With the generalization of digital Television (TV), keeping the channel change delay as low as possible gradually became a difficult requisite in what concerns the resulting user’s Quality-of-Experience (QoE). Frequently, this latency may be higher than 2 s. While many state-of-the-art Set-top-Boxes (STBs) already include a shadow tuner to anticipate the tuning of the next channel, they strive to predict which channel should be pre-tuned, generally opting for one of the adjacent channels. The presented research proposes the use of a predictive system to assist the STB in the forecast of the channel(s) the user will select next. The implemented predictor is based on a Recurrent Neural Network (RNN) and makes use of STB log data concerning the user’s channel changes history to train (and adjust) the model every week. To attain this objective, the most convenient hyperparameter combination that not only fulfilled the aimed prediction accuracy but also suited the rather limited computational constraints of most current STBs had to be identified. The obtained experimental results, validated using four embedded processor families commonly equipping commercial STBs, showed a prediction accuracy of 50.2% for a single-channel prediction and 67.7% when five channels were simultaneously predicted. When combined with the existing dual-tuning system of current STBs, the proposed predictor can save as much as 1000 s per month in TV channel change delays, greatly improving the resulting user’s QoE.