In this paper we present a stochastic variational integrator and its linearization. In order to motivate the use of the proposed variational integrator the Stochastic Differential Dynamical Programming (S-DDP) algorithm is considered as a benchmark for comparison. Specifically, we are interested in investigating if it is advantageous to utilize the variational integrator to propagate system trajectories and linearize system dynamics. Through numerical experiments we show that the Stochastic Differential Dynamical Programming algorithm becomes less dependent on the discretization time step and more predictable when it utilizes the proposed integrator. Furthermore, we show that a significant reduction in computational time can be achieved without sacrificing algorithm performance. Therefore, the proposed variational integrator can be used to enable real-time implementation of nonlinear optimal control algorithms.
Stochastic Variational Integrators and Structured LinearizationThis section presents a stochastic variational integrator used to obtain accurate discretized trajectories of stochastic dynamical Hamiltonian systems represented as d ∂L(q,q) ∂q = ∂L(q,q) ∂q dt + F 0 (q,q, u)dt + where L(q,q) is the mechanical system's Lagrangian given as the difference between the system's kinetic energy, T (q,q), and potential energy, V (q), such that L(q,q) = T (q,q) − V (q), q is the state configuration vector,q is the time derivative of the state configuration vector, dω i , i = 1, . . . , m are Brownian noises, F i (q, u), i = 1, . . . , m are the diffusion vector fields, F 0 (q,q, u) is the forcing function representing deterministic non-conservative external forces, and indicates the stochastic integral is evaluated in the Stratonovich sense Kloeden & Platen (1992). Note equation (1.1) provides the fundamental characteristics of a dynamical system and describes how the system configuration vector propagates through time. However, it does not provide any method or algorithm that can be used to solve for the trajectory of the system. Simply using numerical integration schemes developed for general second order stochastic differential equations will result in numerical errors since the system's fundamental characteristics are ignored. A variational integrator computes a discretized trajectory q = {q 0 , . . . , q N } that approximates the systems trajectory q(t) such that q k ≈ q(t k ), where t 0 = 0, t N = t f and t k+1 − t k = ∆t. The derivation of variational integrators relies on the discretization of the classical Hamilton's variational principle and, in forced systems, the Lagrange-d'Alembert principle. As a result, variational integrators are able to ensure (or strongly enforce) the conservation of fundamental quantities such as momentum and energy Marsden & West (2001). Furthermore, since the derivation IMA Conference on Mathematics of Robotics 9 -11 September 2015, St Anne's College, University of Oxford
Stochastic Variational IntegratorsTo being our review, the discrete Lagrangian, L d (q k , q k+1 ),...