Learning Agent-based Model Predictive Controllers for Holistic Vehicle Control

Abstract

Holistic vehicle control (HVC) is an advanced, integrated approach that optimizes a vehicle’s mobility, stability, and safety by coordinating all available control systems. As the automotive industry progresses towards electrification and increasing intelligence, functional integration has emerged as a dominant trend in vehicle control systems. This evolution necessitates the simultaneous coordination of multiple controllers to achieve diverse objectives. The growing demand for flexibility and reliability in automotive systems has given rise to a “plug-and-play” paradigm in control system design. This approach, while beneficial, poses significant challenges for traditional “all-in-one” integrated control methods, such as integrated model predictive control (MPC).

Distributed control schemes have demonstrated greater scalability and robustness compared to integrated schemes, especially in applications such as vehicle control systems, where managing complex, multi-agent dynamics is important. A recently proposed prominent approach within this framework is agent-based model predictive control (AMPC), where controllers are treated as interactive agents and are coordinated together to achieve a commonobjective iteratively, taking advantage of the distributed control structure. However, in practice, two critical pain points arise:

a) Uncertain contributions from unknown controllers: The optimal control performance from the AMPC highly depends on the prediction accuracy, which requires all agents or their contributions to be accurately known. This requirement is often too idealistic for practical implementation. For example, when a third party develops a “black-box” controller, its underlying algorithm is unknown, and there is no specified interface to know its contribution to vehicle dynamics. This lack of information is likely to cause a significant error in predicting vehicle behaviour, leading to unexpected or even harmful control results.

b) Limitations on controller-oriented decomposition: Decomposing from the objective’s perspective is a more practical and ideal approach in the development of function-oriented or feature-oriented automotive control systems. However, AMPC cannot decompose coupled objectives with shared agents because it is designed only to decompose the integrated system from the controller’s perspective, not from the objectives. The objective in AMPC is usually implemented as a weighted sum of multiple goals, which greatly limits the flexibility of control system design.

This thesis is hence motivated to overcome these two pain points through practical solutions using data-driven machine-learning techniques and flexible distributed schemes for multi-agent-multi-objective (MAMO) control systems.

For the first pain point, this thesis proposes a practical hybrid control scheme: learning agent-based MPC (L-AMPC). This scheme combines the model-based AMPC approach with data-driven learning methods to improve the control performance for multi-agent systems. The Gaussian process regression (GPR) enhanced by an online data management strategy serves as the learning core to predict unknown contributions along the prediction horizon, completing the system model in the MPC for more accurate control. Meanwhile, a stochastic framework is formulated to guarantee control safety and feasibility using soft chance constraints based on the prediction variance. The proposed hybrid control scheme is efficient for real-time implementation and is flexible to any control agent topology.

For the second pain point, this thesis proposes a distributed control scheme: multiobjective AMPC (MO-AMPC). This scheme adapts the alternating direction method of multipliers (ADMM) into a general control strategy that achieves global optimization while decoupling objectives. Three formulations that can maintain convergence while addressing control regularization and inequality constraints are systematically developed. The convergence and computational efficiency of the proposed methods are verified and compared on two vehicle control scenarios with multi-objective configurations.

Furthermore, this thesis proposes a data-driven distributed control scheme based on the MO-AMPC: learning multi-objective AMPC (L-MO-AMPC). To accelerate the converging process of MO-AMPC, a learning-based initialization method for iterations is proposed. The proposed scheme is compared with the MO-AMPC scheme through a path-tracking simulation using various controllers. The results show that the L-MO-AMPC scheme achieves similar control performance while significantly reducing computational time.

All proposed controllers (L-AMPC, MO-AMPC and L-MO-AMPC) are verified by real-time simulations respectively. In addition, the effectiveness and performance of L-AMPC and MO-AMPC are also verified in real vehicle experiments. The results from simulations and experiments could be concluded as follows:

• Compared to AMPC, the proposed L-AMPC can achieve higher tracking performance in well-learned scenarios with the learning capability and always guarantee constraint satisfaction even in less-learned scenarios.

• To decompose the MAMO system from the objective’s perspective, the proposed MO-AMPC achieves the same global optimum as the corresponding integrated MPC with greater flexibility and can potentially reduce the computational cost.

• Compared to model-based MO-AMPC, the proposed L-MO-AMPC has significantly higher computational efficiency while still retaining the property of converging to the global optimum, making it well-suited for real-time implementation.