Most advanced processes struggle to reduce the production cost under constraints. For this, an iterative optimization method called modifier adaptation has been utilized due to its ability to ensure the necessary conditions of optimality even under model-plant mismatch. However, the optimization performance may be degraded by the disturbance which may significantly change the true optimum. In this study, a feedforward decision maker is designed to deal with disturbances in advance and compensate the limitation of feedback scheme of the conventional modifier adaptation. It is constructed by historical data and deep machine learning, and combined with the modifier adaptation. When disturbances occur, the decision maker provides an initial point close to the true optimum by exploiting the historical data. As the information is accumulated, a better initial point for modifier adaptation is obtained. Constrained optimization of numerical example and run-to-run bioprocess are illustrated to validate the utility of the proposed method. (C) 2017 Elsevier Ltd. All rights reserved.