Analysises on factors which impact the performance of the multi level algorithms have been made, and on the basis of which an immune algorithm with selfadaptive reduction has been proposed for the TSP problems. By using an evolutionary reduction set, the proposed algorithm refines the reduction edges which gradually increase in the number and enhance in the forecasting accuracy. As a result, the probability that the refined algorithm finds the global optimal solution can be improved. Experimental results show that the proposed algorithm can achieve better solutions than other approaches.