Skip to Main Content
In this paper , we present a flexible, modular architecture to combine various link state related measurements and prediction algorithms in order to accurately predict link failure in MANETs, while keeping bandwidth and energy overhead low. Our architecture can automatically adjust to varying environment, measurement availability, and prediction algorithm accuracy by dynamically adjusting and learning prediction parameter values. We tested the proposed architecture with a number of prediction algorithms, on two WLAN based scenarios. Our results show that the benefits from using link failure prediction can be substantial in terms of connection time. Furthermore, the aggregated prediction algorithm adapts to the best individual predictor, without knowing which one will be best in advance. The proposed prediction mechanism is also integrated with an 802.11 testbed and Telcordia's Wireless Mobile IP Network Emulator (WISER) for real-time evaluation.