In cellular OFDMA networks, there exists a fundamental trade-off between the achievable cell capacity and the degree of fairness among the users in the cell. Several scheduling algorithms have been proposed which try to balance this tradeoff. The parameterization of these scheduling algorithms to achieve a certain desired fairness level is non-trivial. We show that the optimal fairness parameter settings depend on the system state, such as the current cell load. Our main contribution is a design of a self-optimizing scheduler architecture which includes a controller element that dynamically adjusts the fairness parameters of the scheduler. We demonstrate that with this design, an operator-defined reference fairness level is maintained in scenarios with fluctuating load and thus cell throughput can be improved. It is applicable for a class of proportional fair scheduling algorithms and can be adapted to other algorithms allowing to influence the fairness level.