Search for: All records

Network-wide signal control optimization is of practical importance to shorten and stabilize travel time, improve productivity, enhance energy consumption efficiency, mitigate congestion, and re­duce vehicle emissions. In this study, a deep learning-empowered distributed control strategy is developed to adaptively optimize network-wide traffic signal control coordination. To simplify the problem formulation and enhance its applicability, the entire traffic system is decomposed into multiple areas, and multilayer perceptron concepts are used to formulate traffic control system operations in each area. The distributed deep learning, velocity-based model predictive control (MPC) strategy is designed to optimize traffic signal coordination. Furthermore, a gain-scheduling control model is developed to linearize each learned nonlinear system around its most recent oper­ating status, and then a distributed MPG controller is applied to the linearized systems. Simulation results demonstrate that the proposed control strategy can effectively reduce travel time by 15.1% compared with fixed-time control plans and by 8.0% compared with a decentralized control plan. This study is the first research effort to integrate the deep learning framework and multiagent MPG to optimize traffic control coordination. Moreover, a sufficient condition is theoretically formu­lated for the bounded-input, bounded-output stability of the closed-loop, large-scale traffic system based on the nonlinear small-gain theorem. 

Multi-flow and multi-channel tra!c steering cannot be accurately optimized using the traditional queueing models due to highly dynamic tra!c and channel characteristics. Data-driven leaning based approaches face the scalability and generalizability challenges as the number of flows and the number of channels increase. We propose a hybrid learning approach to per-flow tra!c steering, where a base-level controller generates a base tra!c splitting policy using the measured channel delays and a simplified queueing model, and the base policy is further enhanced by a reinforcement learning agent to control the actual tra!c steering. Through numerical simulations, we demonstrated the e''ectiveness of the hybrid learning controller as it could keep the delay low while avoiding unnecessary tra!c splitting changes. 

Federated Learning (FL) aims to train a shared model using data and computation power on distributed agents coordinated by a central server. Decentralized FL (DFL) utilizes local model exchange and aggregation between agents to reduce the communication and computation overheads on the central server. However, when agents are mobile, the communication opportunity between agents can be sporadic, largely hindering the convergence and accuracy of DFL. In this paper, we study delay-tolerant model spreading and aggregation enabled by model caching on mobile agents. Each agent stores not only its own model, but also models of agents encountered in the recent past. When two agents meet, they exchange their own models as well as the cached models. Local model aggregation works on all models in the cache. We theoretically analyze the convergence of DFL with cached models, explicitly taking into account the model staleness introduced by caching. We design and compare different model caching algorithms for different DFL and mobility scenarios. We conduct detailed case studies in a vehicular network to systematically investigate the interplay between agent mobility, cache staleness, and model convergence. In our experiments, cached DFL converges quickly, and significantly outperforms DFL without caching. 

This paper studies the resilience of cyberphysical systems under denial-of-service attacks. We develop a novel framework for resilient control that avoids the need for detailed information about the system or attacker dynamics by treating the plant–attacker interaction as an interconnected system. Using small-gain analysis and switching systems theory, we derive explicit resilience conditions, and employ reinforcement learning to synthesize an optimal policy directly from input–state data, estimating the required small-gain bounds in a data-driven manner. A numerical example illustrates the effectiveness of the proposed approach.