Self-Organizing Wireless Sensor Networks Solving the Coverage Problem: Game-Theoretic Learning Automata and Cellular...

In this paper, we focus on developing self-organizing algorithms aimed at solving, in a distributed way, the coverage problem in Wireless Sensor Networks (WSNs). For this purpose, we apply a game-theoretical framework based on an application of a variant of the Spatial Prisoner's Dilemma game. The framework is used to build a multi-agent system, where agent-players in the process of iterated games tend to achieve a Nash equilibrium, providing them the possible maximal values of payoffs. A reached equilibrium corresponds to a global solution for the coverage problem represented by the following two objectives: coverage and the corresponding number of sensors that need to be turned on. A multi-agent system using the game-theoretic framework assumes the creation of a graph model of WSNs and the further interpretation of nodes of the WSN graph as agents participating in iterated games. We use the following two types of reinforcement learning machines as agents: Learning Automata (LA) and Cellular Automata (CA). The main novelty of the paper is the development of a specialized reinforcement learning machine based on the application of (ϵ,h)-learning automata. As the second model of an agent, we use the adaptive CA that we recently proposed. While both agent models operate in discrete time, they differ in the way they store and use available information. LA-based agents store in their memories the current information obtained in the last h-time steps and only use this information to make a decision in the next time step. CA-based agents only retain information from the last time step. To make a decision in the next time step, they participate in local evolutionary competitions that determine their subsequent actions. We show that agent-players reaching the Nash equilibria corresponds to the system achieving a global optimization criterion related to the coverage problem, in a fully distributed way, without the agents' knowledge of the global optimization criterion and without any central coordinator. We perform an extensive experimental study of both models and show that the proposed learning automata-based model significantly outperforms the cellular automata-based model.