Numerical optimization problems arise from every field of engineering, science, and business. To tackle the increasingly complex real world optimization problems, effective and efficient algorithms are always on active demand. However, the performance of many existing algorithms deteriorates rapidly as the dimensionality of the search space increases. One natural way to address this problem is to adopt the divide and conquer strategy. In its basic form, the algorithm partitions the high dimensional problem into low dimensional subproblems and optimizes them cooperatively. In this paper, we propose a cooperative particle swarm optimization algorithm for large scale numerical optimization. Firstly, a statistical learning model is proposed to assess the degree of interdependencies among variables. Secondly, based on the variable interdependencies, a method is proposed to decompose a large scale problem into overlapping small scale sub-problems. Finally, a cooperative particle swarm optimization framework is proposed to optimize the sub-problems cooperatively. To examine the performance of the proposed algorithm, experiments were conducted on 10 benchmark functions of diverse complexities. The experimental results show that the proposed algorithm is able to find the nearoptimal solution in most cases, which demonstrate effectiveness of the proposed algorithm.