5. Conclusion
The difficulty of predicting the TCP throughput in interference prone WiFi environments is challenging because of different unpredictable effects such as interference, multipath or other users traffic leading to collisions and unpredictable available capacity. In this paper, we propose to model measured TCP throughput samples as a time series and apply the meta heuristic genetic algorithm to match a set of mathematical functions to best represent the time series. By using the set of functions one can predict future samples, given the GA is trained properly. Using our strategy, one can effectively predict TCP throughput evolution over time by just looking at measured throughput samples without the need to have information available from the TCP stack such as estimates on e.g. round-trip-time or packet loss. We have evaluated the impact of different fitness functions and selection algorithms on the accuracy of predictions. When a more accurate prediction is needed, different retraining schemes can be applied at the expense of more computational power required to find the best set of matching functions. Finally, we have demonstrated that the use of feedback strategies always increases the accuracy of the prediction. In order to improve the accuracy even more, a good strategy has to be found that determines when retraining should be applied. As a future work, we intend to develop heuristics that guide when a retraining should be executed, for example based on knowledge about the TCP congestion control phase. Also, we want to study the impact of different sampling intervals on prediction quality as well as study more scenarios such as different interference situations, different bottleneck links, etc. and their impact on the quality of the prediction.
