I. Introduction
Big data is believed to play a key role in the future transformation of industries [1], [2]. So far, extensive research has been carried out to develop data-driven versions of applied technologies in industrial processes, such as process monitoring [3], [4]. When it is difficult to establish an accurate model for the system of interest through first principles or system identification, or difficult to design a controller based on the model, such as in the presence of highly complex dynamics, data-driven control can be favorable since it avoids the modeling procedure [5]. However, most state-of-the-art data-driven control strategies still involve a model identification scheme and are hence not truly model-free (see, e.g., the review in [6]). Another line of research focuses on extending approximate dynamic programming (ADP) approaches, widely applied in Markov decision processes, to control systems with continuous states and control actions [7], [8]. Despite its model-free property, ADP usually requires the availability of all the state variables, which may be unrealistic. With these considerations, we are motivated to develop a model-free, input-output data-driven control strategy that is applicable to systems with complex dynamics.