Towards automatic generation of congestion control algorithms by coevolving the environment
Teruto Endo[1], Hirotake Abe[1]and Mizuki Oka[1] [1]University of Tsukuba, Tsukuba, Ibaraki 305-8577 Japan enteru.2020@websci.cs.tsukuba.ac.jp
We discuss the applicability of a co-evolutionary algorithmof environments and agents for the automatic generation ofnetwork congestion control algorithms. To co-evolve the network simulation as an environment and the network congestion control algorithm as an agent, we investigated methods for controlling the difficulty of an environment and for generating and optimizing congestion control algorithms, which are necessary for co-evolution. First, we examined the possibility of controlling the difficulty level by varying the amount of cross-traffic generation and the rate of packet loss, which causes throughput degradation. Next, we tested the feasibility of using grammatical evolution for the automatic generation and optimization of congestion control algorithms. The results of these experiments demonstrated that the difficulty of the environment can be controlled by varying the amount of cross-traffic generation and the rate of packet loss. Additionally, it was confirmed that grammatical evolution can be used to optimize the network congestion control algorithm for environments with different parameter settings. In particular, the network congestion control algorithm that we obtained in an environment with high packet loss rates worked robustly in other environments. We show that a co-evolutionary algorithm of environments and agents can be used for the network congestion control algorithm