数学中国w, madio,net The Most Expensive is Not the Best 165 We can do the sensitivity analysis to the weights following the process low. because the changing interval of w and w, is 0, t,+w, if wr and w, are outside the interval, the change won't affect the final order of the two countries; the metrics r and s are insensitive. If not, this change may affect the final order of the two countries If w,>tr, so that the weight of metric r is bigger than wr, the final order of the two countries will be changed; then r is a weight insensitive metric for the country with the lower score. If wr wr, when the weight of metric r is smaller than wr, the finalorder of the two countries will be changed too; then s is a weight-insensitive metric for the country with the lower score. Analysis of American Healthcare System Based on Neural Networks The design of the Back Propagation Network Because of the difficulty of data collection, we choose just satisfaction and seven other indexes as the inputs to a back propagation(BP)neural network: health expenditure per capita, number of doctors per thousand people, number of sickbeds per thousand people. anticipated lifespan, infant mortality, proportion of the healthcare cost in GDP, and extent of healthcare coverage. So, the network should have 7 nerve cells in input layer, and 15(2 x 7+1) nerve cells in middle layer We choose satisfaction to be the target of the network, so there is just one nerve cell in the output layer. According the principles for designing a BP network, the passing function to the middle layer is a sigmoid function. We created the neural network in matlab Application of the BP Network To check the effect on satisfaction when an index changes, we mak e one of them rise by 20% once an nd keep the others unchanged. Doing that, we can get the satisfaction for each year as shown in Figure 1数学中国 www.madio.net