TCP performance in wireless multi-hop networks (WMNs) is hard to achieve due to losses on the wireless channel, interferences and limited resources at individual nodes. Recent research has proposed a simple neural network (NN) structure with one input layer, two hidden layers, and one output layer that efficiently applies congestion control and that results in significant performance improvements compared to conventional TCP variants .
Further, NeuroEvolution of Augmenting Topologies (NEAT) is a method based on evolutionary algorithms that can outperform fixed-topology NNs in reinforcement learning tasks. We expect that NEAT may improve the performance of manually crafted NNs like iTCP even further.
The goal of this project is to assess the ability of NEAT to further improve the performance of an iTCP-based congestion control algorithm in the context of WMNs. The project main goals are:
- Implement iTCP in a network simulation environment (ns-3)
- Use NEAT to generate a modified NN structure for congestion control
- Compare the performance of the modified congestion control to the initial iTCP-based version
 A. B. M. Alim Al Islam and Vijay Raghunathan, “iTCP: an intelligent TCP with neural network based end-to-end congestion control for ad-hoc multi-hop wireless mesh networks”, Wireless Networks, Volume 21, Issue 2, pp. 581–610, February 2015. doi: 10.1007/s11276-014-0799-6
 Kenneth O. Stanley and Risto Miikkulainen, “Evolving Neural Networks through Augmenting Topologies”, Evolutionary Computation 10:2, pp. 99-127, MIT Press, 2002. doi: 10.1162/106365602320169811