Ivanov D.V.

Wireless networks which are getting more wide usage are qualitatively differ from wired fixed networks. They are highly dynamic from their nature and thus require specialized routing protocols to function properly. The efficiency of a routing protocol depends also on the environment conditions. These factors force designers of wireless routing protocols to build protocols with the ability in mind to adjust the protocol to various working conditions defining a set of tunable parameters. However there is no methodology to optimize such routing protocols to given working conditions except manual selection or an expert-s opinion. In this paper a new optimization method for a wireless routing protocol based on a genetic algorithm is proposed. Genetic algorithms are one of the stochastic optimization methods for NP-full problems. To design and test the proposed optimization method OLSR routing protocol was chosen as one of the widely usable routing protocols for wireless networks. The protocol defines approximately 20 parameters which allows to adjust it to given working conditions. The goal is to improve two quality of service criteria: average packet delay and packet drop ratio for the network as a whole. A classical genetic algorithm was used for optimization process. To evaluate chromosomes in a population a simulation for the test scenarios is executed with the protocol parameters encoded in the chromosomes. The optimization and assessment of OLSR protocol was conducted in two network working modes: mobile where nodes moves at a fast pace within an area of the given size, and stationary where nodes move at a very low speed. The results show that the protocol optimized with the proposed method to given working conditions performs well on statistically similar operating conditions. Also, it is shown that in order to achieve robustness to various working scenarios it is necessary to train the protocol on more variable datasets that reflect a broad range of operating conditions. It allows to avoid the effect of overfitting though at the same time the protocol becomes less well tuned to specific conditions.