Much attention has been focused on the problem of mesh generation because of its role as a bottleneck in the finite element analysis. One solution is to develop an efficient adaptive scheme. In adaptive mesh generation, accurate error estmation is indispensable, but it is also important to reflect the calculated result of the estimation exactly in the next step of refinement. In this paper, a new automatic adaptive mesh generation scheme based upon the technique of arbitrarily distributed random numbers is proposed. The steps outlining this scheme are as follows : (1) calculate the distribution of the desired node density from error estimation for a previous finite element solution (2) generate nodes according to the calculated node density distribution (3) construct elements by connecting the nodes. The generation of random numbers according to a given distribution is widely used in Monte Carlo simulation and this technique is applied to generate nodes of the desired density. The applicability of this scheme is examined in some examples. The decrease of the relative error obtained by iterative adaptive remeshing shows the superiority of the proposed method