This paper proposes a method to dynamically control a buffer that determines ground delay times in order to improve the efficiency of air traffic flow management. When arrival demand at an airport exceeds the capacity of the airport, airborne delay occurs in the airspace around the airport. To resolve the imbalance between demand and capacity at the airport, flights that are expected to be airborne-delayed are given a ground delay before departure. A buffer is used to calculate the ground delay time and the buffer is basically fixed in the current operation. We developed a dynamic buffer algorithm based on deep reinforcement learning to improve operational efficiency by changing the buffer based on the uncertainty characteristics of each flight. Numerical simulations were performed in an environment consisting of multiple routes with different uncertainties. The simulation results showed that compared with the conventional fixed buffer algorithm, the proposed method could improve the operational efficiency in terms of airborne delay and throughput loss while balancing them.

When estimating parameters in time-varying systems using data that exhibit sharp changes, it is desirable to place greater emphasis on newer data than on older data. The sliding window method is an effective approach that reduces the influence of older data by discarding it as new data arrive. This allows parameter estimation to be performed using only the most recent data within a fixed-size window. This study considers a recursive algorithm for the least squares estimation with a sliding window. The least squares estimate with the sliding window can be viewed as an estimate with updating and downdating of data. A new, efficient, and numerically stable recursive algorithm is proposed using QR factorization for the updating step and J-orthogonal QR factorization for the downdating step.

This note tackles visual-feedback controller design for a cart-driven inverted pendulum using non-calibrated images taken through a fisheye lens. The images have distortion, which can be modeled as an uncertainty in the output matrix for the measurement signals, due to the use of hemispherical lenses. This problem can be reduced to robust controller design for LTI parametrically dependent systems which belong to a subclass of Linear Parameter-Varying (LPV) systems. This research topic, i.e. robust controller design, is slightly obsolete; however, it is still a challenging problem since the well-established change-of-variables procedure for parametrically dependent plant systems produces not constant controllers but parametrically dependent controllers. On this issue, we first introduce a fictitious parameter vector in the change-of-variables procedure to obtain constant state-space matrices of the to-be-designed controllers in exchange for consequently introduced bilinear terms of decision matrices, and then apply overbounding approach based on Elimination lemma with an auxiliary decision matrix introduced for the bilinear terms to derive our proposed method. The method is still formulated in terms of bilinear matrix inequality, but has no structural constraints for the original decision matrices in the initial feasible solution search. We apply the proposed method to the stabilization problem of the cart-driven inverted pendulum. We successfully design a robust visual-feedback stabilizing controller, and lab-level experiments confirm that our controller stabilizes the system without any calibrations for the distorted images.

Estimated Response Iterative Tuning (ERIT) has been proposed, which can predict input and output data in advance using only a set of closed loops. On the other hand, a reference governor, which is a mechanism for shaping the reference signal, has been proposed as a method to achieve the target response when it is difficult to readjust the controller. However, this method is difficult to apply to systems that are difficult to model, such as large-scale systems. Therefore, this paper proposes a design method for a reference generator that does not require a model, utilizing predicted data calculated by ERIT.