A DBN approach for network availability prediction

Realistic data transfer scheduling with uncertainty

“…Our own prediction work [2,3] shows that even with a good prediction model, average prediction error is in the vicinity of 20%, which is significant. For example, we may predict that Wi-Fi will be available 60% of the time, but it might as well be 40% available or 80% in reality, which are called different availability scenarios.…”

“…We observed in [2,3] that the predictions cannot be done for arbitrarily longer periods with good accuracy. Predictions of network availability can be achieved within acceptable accuracy limits when the durations concerned are short (i.e.…”

“…We addressed that in our previous work and proposed several prediction models [2,3]. With that experience and from others work [29], we make two key observations which need to be addressed in the scheduling, and are described in the first two sections below.…”

“…Fifteen days worth of data which included Wi-Fi presence and GSM presence recorded for every half a minute were gathered from each volunteer [2,3]. GSM was almost always available for all the users as opposed to Wi-Fi.…”

“…We predicted the Wi-Fi availability for a 5 min duration using all the previous days' data plus data of the same day just before the prediction time with a learning process, and we did this repeatedly for all the 5 min blocks in a day, for the last 5 days of each user. Interested readers are referred to [2,3] for more details.…”