The objective for Call admission control (CAC) is to accept or reject request calls so as to maximize the expected revenue over an infinite time period and maintain the predefined QoS constraints. This is a non-linear constraint optimization problem. This paper analyses the difficulties when handling QoS constraints in the CAC domain, and implements two constraint handling methods that cooperate with a NeuroEvolution algorithm called NEAT to learn CAC policies. The two methods are superiority of feasible points and static penalty functions. The simulation results are compared based on two evolution parameters: the ratio of feasible policies, and the ratio of 'all accept' policies. Some researchers argue that superiority of feasible points may fail when the feasible region is quite small compared with the whole search space, however the speciation and complexification features of NEAT makes it a very competitive method even in such cases.