Large reaction networks pose difficulties in simulation and control when computation time is restricted. We present a novel approach to simplification of reaction networks that formulates the model reduction problem as an optimization problem and solves it using a genetic algorithm (GA). Two formulations of kinetic model reduction and their encodings are considered, one involving the elimination of reactions and the other the elimination of species. The GA approach is applied to reduce an 18-reaction, 10-species network, and the quality of solutions returned is evaluated by comparison with global solutions found using complete enumeration. The two formulations are also solved for a 32-reaction, 18-species network.