A Multi-Tier Model for BER Prediction over Wireless Residual Channels

Bit-error rate (BER) modeling and prediction over residual wireless channels, which represent errors not corrected by the physical layer, has emerged as an active research area. Recently, it has been shown that signal to noise ratio (SNR) is a useful side-information that could be employed for BER prediction. In this paper, we propose a novel and accurate three-tier model that leverages a received packet's SNR and checksum side-information to predict BER in future packets over a wireless residual channel. We first observe that direct inference of BER from SNR results in optimistic estimates because of the relatively large amounts of error-free data (in comparison with corrupted data) received on viable wireless networks. Consequently, we propose a model that separates packet-and bit-error prediction. At the first tier, we employ a high-order packet-level Markov model which predicts whether or not a packet is in error. The second tier model is invoked only when a corrupted packet is predicted. The second tier consists of conditional probabilities that predict future SNR values based on the current packet's SNR. Once the SNR is predicted, a third-tier provides the BER estimate for that SNR using a binary-symmetric channel model. We use 802.11b traces collected over an operational 802.11b LAN to compare the performance of the proposed predictor with state-of-the-art predictors. We show that at all three 802.11b data rates (2, 5.5 and 11 Mbps) the proposed model has higher BER prediction accuracy than the optimum Yule-Walker and finite-state Markov chain predictors.