M4: Efficient and Accurate State Estimation and Feedback Control under Uncertainties

该项目基于从M2学习的模型和从M3优化的前馈控制轨迹，专注于对分布式过程的高质量估计和控制。我们假设状态是隐藏的或只有部分可用，因此我们必须估计对分布式过程状态的信念，同时系统地考虑观测中的不确定性。将开发基于RL和MPC的方法来提高受控过程的动力学(稳定性、速度、衰减、精度)，并使过程能够更有目的地进行指导。重点是：(I)分布式非线性过程的信念状态的表示，在复杂性和表示能力之间进行可调折衷；(Ii)在具有可微集合流过滤器的大型分布式状态空间中的随机不确定性传播和过滤的方法，其中状态数可能是数千；(Iii)模块化传感器建模框架，允许在传感器模型之间快速切换，而不需要在成熟的不同阶段重新学习；(Iv)非线性分布式过程的随机反馈控制，基于从M2学习的分布式过程模型，使用基于场景的渐进式随机预测控制来应对模型不确定性和作用于过程的噪声；以及(V)利用给定生产过程的分布式状态和动作空间的基于模型的策略优化技术。研究挑战包括分布式过程模型的高维、对过程的分布式状态和动作表示的利用、状态演化模型中的强非线性、非线性传感器模型以及不同维度的观测。

This project focuses on high-quality estimation and control of the distributed processes, based on the learned models from M2 and the optimised feedforward control trajectories from M3. We assume the state to be hidden or only partially available, so we have to estimate beliefs over the distributed process state while systematically considering uncertainties in observations. Methods based on RL and MPC will be developed for improving the dynamics (stability, speed, attenuation, accuracy) of the controlled process, and will allow the process to be steered more purposively. The focus is on: (i) representation of belief states of distributed nonlinear processes with an adjustable tradeoff between complexity and representation capacity; (ii) methods for stochastic uncertainty propagation and filtering in large, distributed state spaces with differentiable ensemble flow filters, where the number of states may be several thousand; (iii) a modular sensor modelling framework that allows quick switching between sensor models without relearning for the different phases of the maturation; (iv) stochastic feedback control of nonlinear distributed processes, based on the learned distributed process models from M2, with scenario-based progressive stochastic MPC to cope with model uncertainties and noise acting upon the process; and (v) model-based policy optimisation techniques exploiting the distributed state and action spaces of the given production process. Research challenges include the high dimensionality of the distributed process models, the need for exploitation of distributed state and action representations of the process, strong nonlinearities in the state evolution models, nonlinear sensor models, and observations of disparate dimensionalities.