In-flight icing occurs on wings and engine inlets when aircraft fly through clouds containing supercooled water droplets， which is detrimental to flight safety. Hot air anti-icing is one of the most widely used anti-icing technologies. To ensure that anti-icing requirements are met with lower bleed air quantities， an optimization design method based on bleed air control for hot air anti-icing system is developed. A fast prediction method for hot air anti-icing performance based on Proper Orthogonal Decomposition and Radial Basis Function neural networks is constructed， enabling rapid evaluation of the surface temperature and runback water mass flow rate distributions for different anti-icing configurations within the anti-icing optimization design space. Genetic algorithms and Nondominated Sorting Genetic Algorithms are employed for the single- and multi-objective optimization problems of the hot air anti-icing system. The fast prediction method for hot air anti-icing performance is used to evaluate the objective and constraint functions of individuals in the population. While satisfying the constraint functions of the optimization problem， the single-objective optimized design can reduce the required bleed air quantity by 23.41% compared to the baseline design. The multi-objective optimization design produces a series of optimized anti-icing configurations， and the design engineers can choose the most suitable anti-icing configuration design by evaluating the weights of both bleed air reduction and anti-icing performance enhancement.