This article develops a new deep learning framework for general nonlinear filtering. Our main contribution is to present a computationally feasible procedure. The proposed algorithms have the capability of dealing with challenging (infinitely dimensional) filtering problems involving diffusions with randomly-varying switching. First, we convert it to a problem in a finite-dimensional setting by approximating the optimal weights of a neural network. Then, we construct a stochastic gradient-type procedure to approximate the neural network weight parameters, and develop another recursion for adaptively approximating the optimal learning rate. The convergence of the combined approximation algorithms is obtained using stochastic averaging and martingale methods under suitable conditions. Robustness analysis of the approximation to the network parameters with the adaptive learning rate is also dealt with. We demonstrate the efficiency of the algorithm using highly nonlinear dynamic system examples.
This work focuses on optimal pollution controls. The main effort is devoted to obtaining approximation methods for optimal pollution control. To take into consideration of random environment and other random factors, the control system is formulated as a controlled switching diffusion. Markov chain approximation techniques are used to design the computational schemes. Convergence of the algorithms are obtained. To demonstrate, numerical experimental results are presented. A particular feature is that computation using real data sets is provided.
This letter develops a new approach for a class of stochastic control problems under partial observations. The new approach is a deep filtering based method. Algorithms are proposed, and comparisons to the Kalman filter are provided. Both quadratic and non-quadratic costs are considered. Numerical experiments and production planning case studies are presented to demonstrate that training with a constant-gain linear feedback integrated with the DF state estimator can meet the needs of a broad class of optimal control problems under partially observation of the state.