Abstract:
This dissertation aims at developing a mathematical framework by using an automated learning to help in the decision making in a route selection process of wireless sensor network in flood warning application. The automated learning is the reinforcement learning (RL) where the best possible route will be determined from the policy (set of actions) obtained from the action-value function by considering on the small-scale network scenario with a static topology and the large-scale network scenario with a dynamic topology. The application of the monte Carlo algorithm in flood warning application has been used to determine the best possible route selection with three constraints which are remaining energy, remaining lifetime and path reputation. These three constraints have been considered as a reward function in small-scale network scenario. By using path reputation, this technique can be alleviated the single point of failure of the system. The preliminary result confirms that the monte Carlo algorithm performs well and effective. Therefore, the investigation on a large-scale network has been considered. The experimental settings for a large-scale network begin with the consolidation of Monte Carlo algorithm by changing state variable from simple to the advance. Previously, the state variable has not incorporated the environmental status but it will be included in a large-scale network. The main idea is to extend the proposed algorithm towards the actual environment scenarios. Thus, the effect of reward function to the solution of monte Carlo algorithm has been observed as well as the optimal load balancing. Only remaining energy and its reputation are required for being state variables. From the investigations, the reward function with three components (remaining energy, remaining lifetime and the path reputation value) performs well. Note that there are many possible weighted functions can be used but the results in terms of total remaining energy from all sensor nodes are slightly different. Finally, the weight has been chosen equally. The performance measurement has been done in terms of the ability to maintain the network connectivity by balancing and sharing the traffic load to alternative routes. The benchmarking methods are shortest path, max-min method, uniform random, monte Carlo with non-intelligence and monte Carlo with intelligence. The results show that the proposed method can guarantee the link connectivity time longer than the worst case method up to 20% (the shortest path). Finally, the computational complexity is considered as well as the possibility work towards future.