Since the resources stored in the cloud is huge and the cost of each task in cloud resources is different, the simple Round Robin algorithm and FIFO algorithm for task scheduling can't met the growing scale of cloud computing. The traditional GA algorithm for task scheduling, which has the defect of premature convergence, only takes the time cost into consideration, but ignores the consumption of resources. In order to solve the problems exists in multi-task scheduling in cloud computing mentioned above, we propose an improved GEP algorithm with double fitness functions (DF-GEP), and also constructs a new ETCC matrix which not only considers the running time of all tasks, but also takes the running cost of the tasks into consideration. This improved algorithm reduces the optimization time, and falls into the local optimal solution hardly at the same time. This improved algorithm expresses a good convergence, through experiments compared with GA and ordinary GEP algorithm by using the Map/Reduce programming model.