In this project, we intend to proposes a generic multi-objective AMPC scheme for multi-agent systems. Different formulations tailored from the ADMM are proposed systematically. The global convergence of the proposed formulations was also proved. Furthermore, the proposed scheme with optimal parameters was also superior to integrated MPC on computational efficiency in real-time experiments.
In practice, control systems often include multiple objectives, where coupling between agent groups affects different objectives, posing a big challenge for distributed control schemes. A large integrated MPC may not be practical due to the limitations arising from the physical separation of controllers or the limited computational capabilities of some real-time embedded systems. Thus, we proposes the multi-objective agent-based model predictive control (AMPC) scheme as a generic and practical distributed solution for multi-agent control systems.
An essential requirement of the distributed control schemes for multi-agent systems is to obtain the same result as the integrated one: the global optimum. This requirement becomes challenging for many previously proposed distributed MPC (DMPC) solutions when the system cannot be decoupled into fully independent subsystems. To this end, this project systematically proposes a multi-objective agent-based MPC scheme tailoring from the alternating direction method of multipliers (ADMM), in which objectives are distributed with a corresponding controller agent group. It is guaranteed to achieve the global optimum iteratively through information exchange between agent groups. The proposed scheme has great flexibility that objectives (i.e., product features) can be designed in a “plug-and-play” fashion. Besides, the proposed scheme also superior to integrated MPC on computational efficiency in real-time experiments.
Experimental Results in Cruise Control
Two objectives of longitudinal speed tracking (LST) and yaw stability control (yaw rate tracking, YRT) should be met simultaneously in the cruise control. Comparative tests of double line change (DLC) during acceleration and single line change (SLC) during deceleration were conducted on high-friction road surfaces. The error between the desired and measured yaw rates is significantly reduced when the objective of YRT was enabled, indicating that the vehicle’s yaw stability has been improved.
The results of the suboptimality, iterations, and time consumption ratio are shown in the results. The results demonstrate that even for such a small system, the proposed multi-objective AMPC is highly computationally efficient and outperforms the integrated MPC, while is always converged to the global optimum.
Experimental Results in Holistic Stability Control
Vehicle stability control is another challenging task that requires coordinating multiple controllers to meet multiple objectives. In this example, two-level stabilities is considered, where the vehicle stability control (VSC) is to track the desired yaw rate and the wheel stability control (WSC) is to track the desired wheel speed for the maximum tire capability.
It can be seen from the results that the significant longitudinal wheel slip and over-steering happened during the first experiment. The driver needed a large counter-steering maneuver to keep the vehicle stable. However, when both stability objectives were enabled, yaw rate error was significantly reduced and no significantly slip appeared. Also, the proposed multi-objective AMPC is highly computationally efficient and outperforms the integrated MPC, while is always converged to the global optimum.
