Abstract: In a distributed system without a central administrator, suppose all independent processing units do not know the conditions of one another, a new system is designed to undertake the tasks in satisfying proportions to cooperatc for the best results, in which only self-learning is used.This paper also analyzes the performance of this system by Markov chains, and presents a robust method of self-learning for independent processing units in this kind of systems.The method can also be used to solve the general assignment problem.