After reviewing the background and description of cutting and packing problem and packing algorithms, this paper proposes an improved genetic algorithm for non-guillotine rectangles packing problem. First, self-adaptive mutation probability is adopted to avoid pre-maturity, some commonly used crossover operators are compared, and a stochastic hill-climbing operator is designed to improve the local searching ability. Second, the greedy mechanism is introduced to acquire good packing pattern after packing order is determined by the GA. Computational results show the effectiveness of these improvements.