Towards a generalized self-organizing multi-agent system

A Multi-agent system (MAS) is one which has a number of independent software agents interact with each other to achieve a common goal or goals automatically. It is commonly known that many species are living in the form of a MAS as it has high adaptability on surviving in an unstructural environment. Based on the benefits of group living observed from the nature, MAS is a potential direction to solve a wide range of engineering problems including robotics, chemistry, finance and genetics.; In the first part of this thesis, a radial basis function (RBF) network called "Agent Swarm Regression Network (ASRN)" is proposed in which the training algorithm is modeled as an evolution of a rule-based MAS. Three sets of experiments show that the performance of ASRN is better than that of a conventional approach in terms of computation, complexity and memory usage. The experimental results show the acceptable generalization ability and accuracy of ASRN.; In order to maximize the ability of MAS, self-organization is an essential element to be considered. Self-organization refers to a process in which the knowledge of a system accumulated automatically without being guided by an outside source or super-intelligence. In addition, the knowledge is accumulated only when the system interacts with the environment. Therefore, a robust self-organizing MAS should use a minimum number of interaction to construct the strategy for a maximum award. This goal can be achieved by involving the unstructural environment modeling and optimal response generation.; As the goodness of a response is measured from the quantity of award received, optimization is necessary in self-organization. We proposed a novel global optimization algorithm called "Creativity Driven Optimization (CDO)" in the second part of this thesis. By introducing the idea of creativity, CDO requires fewer evaluations than that of three reference methods to search for a global optimum.; To tackle the modeling process of an unstructural environment, a sequentially trained neural network called "Memory Re gression Network (MRN)" is proposed in the third part of this thesis. Based on the human's learning strategy, fewer training samples are required to train MRN in which the accuracy and generalization of MRN is similar to a reference network. After estimating the environment, the optimal response function is constructed by the estimated environment with the cooperation of a newly proposed algorithm: Response Knowledge Learning (RKL). The simulation result shows that the predator trained by RKL catches the prey within 15 steps after 50 independent successful hunting trials.; In the last part of this thesis, we presented a procedure learning algorithm of self-organizing agent (PLSOA) that consists of CDO, MRN and RKL. Instead of searching for the current response with a local maximum award with reinforcement learning, PLSOA generates a response sequence by optimizing an adaptive objective function that can adjust iteratively. The experimental results of three benchmark problems show that PLSOA is able to generate nearly optimal-length response sequences in three benchmark environments. In addition, the proposed algorithm has an advantage over the reference methods in terms of reduction on procedure evaluation.; In this thesis, we have made some major contributions towards a generalized self-organizing MAS which try to mimic the MAS in nature.