Research On The Application Of Multi-Agent System For Optimization Problem

This dissertation mainly researches applications of Multi-Agent system (MAS) for optimization problems, such as function optimization, container stowage problem (CSP) etc. The dissertation also constructs the mathematic model of CSP and gets the optimal solution with COIN model, which can be used to validate the solution of Multi-Agent system for CSP.Agent entity contains knowledge, belief, intent and goal etc, which has the features of autonomy and interaction. It can feel the change of environment and take some reactions to the change. Agent is composed with repository, communication module and choice function component. Generally, Agent is a system based on hardware or software, which has features of autonomy, sociality, reaction and skill. We always use Multi-Agent system solve problem, because of the ability limitation of single Agent. MAS assembles different ability agents as a whole, and use the interactions: division and cooperation between them solve problem. With the development of distributing artificial intelligence, MAS has been used in variety fields.Optimization is the spirit of mathematic model. As a special mathematic filiation, optimal techniques have variety application fields. Function optimization is a classical deputy of optimization problem, the ability of algorithm in solving function optimization stands for the ability in optimization.Agent is an individual that has intelligent behavior, and the individual is given target dissimilarly based on different concrete problems. Agent will change its own attribute according to its belief in the process of iteration to attain the purpose of tending target. We choose the domain of definition of the function as the existence environment of the Agents, and choose the return value as the target of the Agent. Each Agent is given a belief randomly in the beginning, in the period of each iteration of the system, all individuals decide the directions of moving on and the size of step according to its own belief by using the learning machine, then adjust the belief according to the changes of the return value after moving ahead. At last, Agents interact with others in the same scope. After finite iteration, mass of agents can reach the optimization in the system. The belief adjusting function makes the system converge in short time and avoids localoptimization.Container stowage problem (CSP) is a kind of combination optimization problem. Base on the special demands, logical arrangement of containers in the vessel makes the system return an expectant solution; the goal is always shift minimization in CSP. From 1970s, container transport largely increased, now days Over 60% cargoes are transported by container, and even 100% between some developed countries [6]. The transport companies use large container vessel to getting the big scale economy profit, which led to the increasing of shift and the shift is the largest part of transport cost, shift a TEU will cost 400MCY. Even more, shift is also an important factor, which can influence the stowage service time. So, the research goal of CSP is how to reduce shift in the largest degree, consequently, reduce stowage work time and reduce superfluity cost produced by shift. The dissertation proposes a CSP mathematic model. The model can be used to optimize Containers Ship Stowage Problem satisfying restrict conditions, which can minimize the container shifts, with vessel stability. This dissertation uses COIN (Common Optimization Interface for Operations Research) to solve the model in experimentsand get the optimization; the results will be used to validate that the MAS's solution at CSP is valid and acceptant.For solving CSP problem with MAS, we choose the available spaces of the container vessel as the existence environment of the Agents, and choose the system shifts as the target of the Agent, and choose individual shifts as the secondary target of the Agent. Let each Agent stands for a container, then we give Agent a belief and location randomly in the beginning, in the period of each iteration of the system, all individuals decide the directions of moving on and the size of step according to its own belief by using the learning machine, then adjust the belief according to the change of system shifts and individual shifts. The system can get a capable solution using the mechanism after several simulation experiments. Compared with the COIN's optimization in time cost, shift and vessel stability, we can see the efficiency of MAS model in CSP problem application.