A Technique to Enable Online Machine Learning Applications for Simulation Optimization

Simulation optimization refers to an optimization problem with a stochastic and potentially computationally expensive objective function. Machine learning (surrogate modeling) techniques have significant potential for enabling efficient simulation optimization, but typically require the user to retain all input/output pairs evaluated by the objective function and can suffer from numerical stability issues if samples are tightly clustered. A modification to the traditional surrogate modeling process is presented that enables robust online learning through use of a Kalman filter. The results of training kriging and radial basis function models are presented for both the conventional training process and the modified training process. The modification results in minimal loss in the ability of the surrogate model to represent moderately complex objective functions.