Operating large multi-reservoir systems is critical for effective water allocation, hydropower generation, and economic development. However, conventional robust planning-based methods scale poorly beyond single-reservoir or cascaded systems due to computational intractability. Addressing inter-reservoir relation extraction and coordination under conflicting objectives, uncertain inflows, and complex network couplings therefore remains an open challenge. To tackle this problem, we introduce the multi-agent reservoir cooperative operation (MARCO) environment, which integrates multiple objectives, heterogeneous topology, and stochastic inflows in a unified framework. An algorithm is then designed to construct a multi-relational directed acyclic graph (MR-DAG) that encodes the underlying topology through coupled objectives and entity relations. Building on this representation, we propose the multi-agent relational directed acyclic graph transformer (MAR-DAGT), a reinforcement learning algorithm that performs typed message passing for efficient feature extraction and employs acyclic decision-making to exploit causal structure for improved credit assignment. Extensive experiments on MARCO show that MAR-DAGT consistently outperforms other MARL and optimization-based baselines in terms of objective satisfaction and robustness to inflow uncertainty.

In this study, a novel optimal tracker is developed for the linear quadratic tracking (LQT) problem using an output–feedback adaptive dynamic programming (ADP) framework. By leveraging the minimal polynomial of the exosystem matrix, we parameterize the steady-state input, state, and output, and incorporate them into the performance index of the LQT formulation. Unlike existing studies on LQT, in this study, we introduce a high-order filter to make the signals related to the high-order derivatives of the system output and input observable. Subsequently, we develop an incremental data ADP algorithm to learn the optimal dynamic output–feedback tracker, eliminate the impact of high-order filtering error and state reconstruction error on the iterative learning equation (ILE) within a specified time. Finally, the comprehensive simulation results show the effectiveness of the proposed output–feedback tracker.

This study investigates the state estimation problem of multi-agent systems under a homologous sensor attack. A distributed filter is proposed to achieve a minimum variance unbiased (MVU) estimation of system states and attacks in the presence of measurement noise. A gain matrix selection method for implementing the MVU estimation is also provided. The proposed filter can be used for positioning corrections affected by global positioning system (GPS) meaconing attacks. This study treats the positioning offset caused by GPS meaconing attacks as a zero-mean white random variable and verifies the validity of this hypothesis through experiments with real GPS signals. Moreover, this study comprehensively analyses the integration of the filters into practical systems. Finally, the effectiveness of the proposed results is verified using simulation examples. Note to Practitioners–This study introduces a filter that can achieve GPS positioning calibration under meaconing attacks. This filter treats the true positioning of the system as a state and the deviation caused by meaconing attacks as a homologous attack. The filter employs a collaborative filter to reconstruct the state, thereby enabling positioning calibration. Additionally, the study explores the relationship between the homology of meaconing attacks and the estimation error of filters. This result reveals that as the homology of attacks increases, the performance of filters also improves.

This article explores the issue of fault-tolerant optimal pursuit strategies in a nonlinear pursuit-evasion (PE) game involving multiple pursuers and a single evader. The main challenge lies in ensuring the successful capture of the evader despite the presence of actuator partial loss of effectiveness and bias faults within the pursuers group. To overcome this challenge, a two-layer control architecture is proposed. At the control layer, an integral sliding-mode controller is developed to mitigate the impact of bias faults, and an adaptive estimation mechanism is incorporated to identify the fault parameters. At the decision-making layer, performance index functions for both the pursuers and the evader are formulated based on the PE state error and their respective control strategies, and optimal pursuit and evasion strategies are derived by solving the associated Hamilton–Jacobi–Isaacs (HJI) equations. Furthermore, adaptive dynamic programming (ADP) is employed, with each participant using a critic network to approximate the optimal strategies for the pursuers and the evader. The proposed control mechanism is theoretically proven to ensure that all closed-loop signals remain uniformly ultimately bounded, allowing the faulty pursuers to successfully capture the evader. Finally, the effectiveness of the approach is validated through two simulation examples.