Study Of Applications Of Machine Learning On Decision And Controlling In Multi-Agent Systems

The application of machine learning on decision and controlling in Multi-Agent Systems is studied: Supervised Learning algorithms are widely used in Multi-Agent Systems. The Modal Based Learning is presented, witch is beyond any specific algorithm, and is dedicated to solve the problem of how to acquire training data for supervised learning algorithms. For the problem of noise in training data for supervised leaning, the Clustering based Date Preprocessing Algorithm (CDPA) is presented, which can eliminate noisy data effectively and then improve the efficiency of learning algorithms. Potential Role Assignment is introduced to Multi-Agent Systems in order to avoid the possible instability which can be caused by frequent role switching of Agents.Machine learning algorithms can be divided into three categories according to whether there is a teacher, which are supervised learning, unsupervised learning and reinforcement learning. In Multi-Agent Systems, supervised learning algorithms are widely used such as the Artificial Neuron Network and Decision Tree. As a principal of these algorithms, training data are needed for the learning unit to generalize. Modal Based Learning (MBL) is a method that can solve the problem of how to acquire training data. This method is beyond any specific algorithm and utilizes the observability of the system to collect raw online data, then enable the learning unit to carry out offline generalization.Besides the problem of training data acquirement, another problem that almost all supervised learning algorithms confront is noise, which can be cause by variousfactors. Though algorithms with robustness and error toleration can not be affected by noise, as far as final result is concerned, yet the speed of contraction is lowered. As for those algorithms that are highly sensitive to noise, the whole process can be compromised. In this paper, we introduce Cluster based Data Preprocessing Algorithm, which performs clustering operation on original training data twice and will eliminate noise effectively. This algorithm can be carried out in polynomial time and can be adjusted through parameter tuning. Application of this algorithm can result in dramatic improvement of the learning efficiency.Role modal can be defined in some Multi-Agent Systems, robot soccer systems are typical Multi-Agent Systems that have a role modal. In this paper, Potential Role Assignment is introduced, which enhance the traditional role modal by defining potential role value of each Agent, So that the possible instability which can be caused by frequent role switching of Agents can be thoroughly avoided. By adjusting the potential role values dynamically Agents can actually carry out reinforcement learning, and the performance of the whole system can be improved.