Automating the construction of neural models for control purposes using genetic algorithms

With the purpose of automating the process of modelling and improving the quality of the control solution, a strategy based on genetic algorithms for determining the structure of each model has been developed and tested on a real system with measurement noise. The models were produced using feedforward neural networks and were tested in different control loops such as direct inverse control and internal model control and compared with the models obtained using the expertise of a control engineer. Several difficulties are reported as being obstacles to the success of the strategy and the solutions presented. The overtesting problem and a hybrid general training/specialized training solution are the major contributions of this work.