Dynamic Network Selection Algorithm Based On Evolutionary Game

In recent years,802.11-based wireless LAN(WLAN) technology is developing rapidly. Compared to wired local area network (LAN), wireless LAN is widely used with the advantages:low-cost, high-speed, easy to install and easy to operate. A large number of access points (APs) are densely deployed in locations such as airports, hotels, schools to support network usage by many simultaneous users, and thus there are several APs for a wireless device to associate to. Existing program simply determines which AP the wireless devices access to only with the received signal strength (RSSI), which could lead to overload of the APs associated by most of STAs, but the others are idle, and thus exacerbating the load differences between APs, reduce the utilization of network resources, affecting the system performance.A detailed study and analysis of the progress of load balance in WLAN at home and abroad were summarized in this paper. Then, we modeled the network selection problem as the evolutionary game. In this model, STAs in the same overlapping area consitute a population, and all the populations are independent in this paper. STAs as the game playes compete against others in the same population for network resources. Evolutionary equilibrium is the final solution.This paper proposes two dynamic network selection algorithms, including centralized population evolution algorithm and distributed reinforcement-learning algorithm, to alleviate the load imbalance. With population evolution algorithm, STAs compute their payoffs, and send the informations to a Central controller. After that, the controller computes the average payoffs of populations and payoffs provided by APs, then returns these informations to corresponding STAs. Population evolution algorithm combines advantages of AP-dreiven and STA-driven load balance technology, not only converges fast, but also will not has the risk of "island". In contrast, with reinforcement-learning algorithm converges slowly as the STAs have to gradually learn from the network environment and adapt the network selection decision to reach the optimal state. Compare with Population evolution algorithm, reinforcement-learning algorithm does not need extra hardware has low communication overhead, and is beneficial to network expanding.Finally, above-mentioned algorithms are achieved in ns-2. The simulation results demonstrate that both these these algorithms can effectively balance load among APs. Compared with reinforcement-learning algorithm, the population evolution algorithm can reach the evolutionary equilibrium faster.