Mathematical Problems in Engineering

Volume 2018 (2018), Article ID 7908378, 11 pages

https://doi.org/10.1155/2018/7908378

## A High-Precision Single Shooting Method for Solving Hypersensitive Optimal Control Problems

^{1}College of Astronautics, Northwestern Polytechnical University, Xi’an, Shaanxi 710072, China^{2}Department of Aerospace Engineering and Technology, Politecnico di Milano, 20156 Milano, Italy

Correspondence should be addressed to Binfeng Pan; nc.ude.upwn@gnefnibnap

Received 9 December 2017; Revised 27 February 2018; Accepted 12 March 2018; Published 15 April 2018

Academic Editor: Muhammad N. Akram

Copyright © 2018 Binfeng Pan et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

Solving hypersensitive optimal control problems is a long-standing challenge for decades in optimization engineering, mainly due to the possible nonexistence of the optimal solution to meet the required error tolerance under double-precision arithmetic and the hypersensitivity of the optimal solution with respect to the initial conditions. In this paper, a new high-precision single shooting method is presented to address the above two difficulties. Multiple-precision arithmetic and Taylor series method are introduced to provide the accurate optimal solution with arbitrary higher significant digits and arbitrary higher integral accuracy, respectively. Besides, a new modified bidirectional single shooting method is developed, which fully utilizes the three-segment structure of the hypersensitive optimal control problems and provides appropriate initial guess that is close to the optimal solutions. Numerical demonstrations in a typical hypersensitive optimal control problem are presented to illustrate the effectiveness of this new method, which indicates that the accurate optimal solution of this challenging problem can be easily solved by this simple single shooting method within several iterations.

#### 1. Introduction

An optimal control problem and its associated Hamiltonian boundary value problem (HBVP) are called hypersensitive if the time interval of interest is long relative to the rates of expansion and contraction of the linearized Hamiltonian dynamics in certain directions in a neighborhood of the optimal solution. Solving hypersensitive optimal control problems (HOCPs) is a challenging task, which has been studied extensively for decades [1–6].

Typically, there are two types of methods, usually categorized as direct and indirect methods, for solving HOCPs. The direct method converts the optimal control problem into a nonlinear programming problem (NLP) by appropriate discretization [7]. Various direct methods have been developed during the past few decades, such as direct multiple shooting method [8], gradient-based optimization techniques [9, 10], differential inclusion method [11], interior point method [7, 12], collocation method [13, 14], pseudospectral method [4, 15–17], and control parameterization method [18–20]. Direct methods are straightforward and robust to accommodate complex conditions, such as path constraints and intermediate constraints. When applying direct methods to solve HOCPs, the key is to use a higher density of nodes in the first and third segment, according to the three-segment structure characteristic of HOCPs, which is described qualitatively as “take-off,” “cruise,” and “landing” by analogy to an airport-to-airport trajectory for an aircraft [1]. Reference [3] introduced a simple method for mesh point distribution for solving HOCPs based on density functions, and the problem of mesh refinement is subsequently converted to a problem of finding an appropriate density function. Reference [4] presented a hp-adaptive pseudospectral method to solve HOCPs, which iteratively determines the number of segments, the width of each segment, and the polynomial degree required in each segment. Reference [5] proposed a symplectic algorithm with nonuniform grids combining with density functions to solve HOCPs. Later, [21] applied GPOPS-II, a commercial software program that employs a Legendre-Gauss-Radau quadrature orthogonal collocation method to solve HOCPs. In their method, an adaptive mesh refinement method is implemented to determine the number of mesh intervals and the degree of the approximating polynomial within each mesh interval to achieve a specified accuracy. However, there are several disadvantages for solving HOCPs by direct methods. The first one is that the obtained optimal solution by direct methods is merely approximate to the accurate solution. The error between these two solutions is usually small and acceptable for most optimal control problems; however, this error may be extremely large and unacceptable for HOCPs, as demonstrated in this paper later. Besides, direct methods cannot provide insight regarding the optimal solutions [6].

Unlike direct methods, indirect methods transform the original problems to HBVPs by first-order optimality conditions [22], which are helpful to provide the dynamical information of the systems. Besides, the optimal solution is guaranteed to be a least extremal. However, it is extremely difficult to solve HOCPs by indirect methods, mainly due to the high sensitivity to the initial unknowns. Based on the three-segment structure, [1, 2, 23, 24] developed a dichotomic basis method to solve HOCPs. The core idea is the decomposition of contracting and expanding behaviors of the system by coordinate transformation, and the iterative process is required due to the approximation of dichotomic basis. Recently, [6] presented a finite-time Lyapunov analysis based method to construct an approximate dichotomic basis and developed a corresponding manifold-following solution approximation method. However, the optimal solutions obtained by the above methods are only near-optimal, and, as pointed in [6], the sensitivity of the optimal solution to the errors in the unknown boundary conditions may cause ill-conditioning, which potentially exceeds the available precision. Thus, to the best of the authors’ knowledge, the above difficulties are not well solved in the existing literature, and obtaining the accurate optimal solutions of HOCPs by indirect methods is still extremely daunting.

In this paper, a new high-precision single shooting method is presented to address the above two main difficulties for solving HOCPs by indirect methods: the possible nonexistence of the optimal solution to meet the required error tolerance under double-precision arithmetic; the hypersensitivity of the optimal solution with respect to the initial conditions. Multiple-precision arithmetic and Taylor series method, which provide higher significant digits and higher integral accuracy, respectively, are introduced to address the first difficulty. A modified bidirectional single shooting method (MBISSM) is developed to settle the second one, which introduces an intermediate point along the equilibrium segment as the starting point and shoots to both boundary points simultaneously. Finally, numerical simulations are provided to demonstrate the effectiveness and efficiency of the proposed method, which reveals that the challenging HOCPs can be easily solved within several iterations by this proposed method.

#### 2. Hypersensitive Optimal Control Problem

Consider the optimal control problem with the dynamical equations as follows:where and are the state and control vector at time , respectively, and is the smooth function. The performance index iswhere and are the prescribed initial time and terminal time, respectively. The boundary conditions are defined aswhere and are known.

According to the optimal control theory [22], the corresponding Hamiltonian function is implicitly dependent on time, which is expressed aswhere is the costate vector. The Hamiltonian differential equations are as follows:and the optimal control vector satisfies

Denote , as an alternate expression for the Hamiltonian system in (6). For sufficiently large value of , that is, is long relative to the rates of both expansion and contraction in certain directions in the neighborhood of the optimal solution; this problem becomes completely hypersensitive. It is a degenerate class of two time-scale HBVPs that is composed of fast and slow subsystems. Otherwise, the problem is partial hypersensitive [1, 2]. In this paper, the completely hypersensitive case is considered, and the saddle type equilibrium point is supposed to exist. As illustrated in [1, 2], the optimal trajectory shows the three-segment structure geometrically, which is described by where and denote the durations of the initial and terminal segment, respectively. is the optimal solution of the initial segment with initial conditions , is the optimal solution of the terminal segment with terminal conditions , and is the optimal solution of equilibrium segment with . Please notice that and should be large enough to allow the boundary-layer segments to reach with sufficient accuracy [2].

Based on the observation that , [1, 23] presented a dichotomic basis method for HOCPs. In their method, the original optimal control problem is spitted into three subproblems.

*(a) Subproblem A*. Solve Hamiltonian system with the optimal control determined by (7), initial conditions defined in (3), and terminal conditions given as where is the guessing time duration of the initial segment.

*(b) Subproblem B*. Solve Hamiltonian system without control effort, that is,together with the initial conditions and terminal conditions as follows:where is the time duration of the terminal segment. Since the control variables are totally determined, this subproblem is actually an initial value problem solved by a numerical integration method.

*(c) Subproblem C*. Solve Hamiltonian system with the optimal control determined by (7), the initial conditions defined in (4), and terminal conditions given as

Please notice that subproblem C has the same structure as subproblem A, except that the time direction is opposite. Thus it requires a backward integration to complete numerical computation.

Dichotomic basis method is elegant and efficient; however, the obtained optimal solution by this method is only near-optimal. Thus, obtaining the accurate optimal solution of HOCPs by indirect methods is still an open challenge.

#### 3. High-Precision Single Shooting Method

In this section, a new high-precision single shooting method is developed to obtain the accurate optimal solution for HOCPs, which combines with three important themes: multiple-precision arithmetic, Taylor series method, and a MBISSM. In this proposed method, multiple-precision arithmetic and Taylor series method provide the ability of high-precision computation with higher significant digits and higher integral accuracy, respectively, and the MBISSM provides appropriate initial guess that is in the close neighborhood of the optimal solution.

##### 3.1. Multiple-Precision Arithmetic

Currently, most optimization applications and algorithms utilize double-precision (64-bit) accuracy, which allows numerical computations to be accurate to ~15 significant digits and is more than sufficient for most optimal control problems. However, for a rapidly growing body of applications, such as HOCPs considered in this paper, double-precision arithmetic may not be sufficient any more. Therefore, a higher level of numeric precision is required, and multiple-precision arithmetic is an effective way.

Typically, the ability of multiple-precision arithmetic is commonly provided in software, as hardware support for higher precision is severely lacking. It is highly likely that a commodity processor supporting higher floating-point precision will be developed in the near future [25]. Quadruple (128-bit or ~30 significant digits) or octal (256-bit or ~60 significant digits) precision arithmetic is a typical choice. For some specific problems, hundreds or even more significant digits are required to obtain numerically meaningful results.

Nowadays there are several commercial or freely available multiple-precision software packages, which usually provide high-level language interfaces, custom data types, and operator overload features. In this paper, a Multiprecision Computing Toolbox developed by Advanpix [26] is applied for high-precision computation. This software is a MATLAB extension for computing with arbitrary precision; that is, the toolbox equips MATLAB with a new multiple-precision numeric type and extensive set of mathematical functions that are capable of computing with arbitrary precision. Thus many existing MATLAB programs can be directly converted to achieve arbitrary precision with small efforts.

##### 3.2. Taylor Series Method

Numerical methods for integrating ordinary differential equations have been studied since the end of the last century, and a large number of integration formulas have been developed. Typically, most numerical integration methods are designed to approximate the exact solution of the ordinary differential equations, which gives rise to truncation errors. Generally, the truncation errors remain to be sufficiently small for most optimal control problems with common numerical integration methods, such as the well-known Runge-Kutta methods. However, for HOCPs considered in this paper, the truncation errors would be amplified to be incredibly large with these methods, even when multiple-precision arithmetic is applied.

The truncation errors can be reduced by two different ways: by reducing the step size and by using the higher-order integration formula. If the step size between two adjacent values becomes smaller, the truncation errors of the numerical integrations would decay. The disadvantage is that the step size may decrease to be so small that it may take extremely long time to complete the integration, and the round-off error would be increased [27]. Thus, a higher-order integration formula, called Taylor series method [28], is preferred to provide accurate numerical integration solution. Taylor series method is one of the oldest methods, which traces back to Newton and Euler. It has an advantage that its formula at an arbitrarily high order can be easily expressed in the united form with large step sizes and thus small values of computing time is required [29]. Therefore, from the viewpoint of numerical simulations, it is rather easy and attractive to use this method at a very high order so as to reduce the truncation errors to a required level.

Consider the initial value problem as follows:where , is the specified initial time, is the prescribed initial conditions, and is the smooth function. According to the rules of Taylor series method, the value of the solution at is approximated by the th degree Taylor series of developed at and evaluated at , that is,where is the step size and is defined as

In this paper, a variable step size and variable order (VSVO) scheme, which aims to control the integration error automatically by adjusting the integration step size through the criteria related to the prescribed tolerance and variable order formulation, is implemented during the integration by Taylor series method. The details of VSVO scheme can be found in [30].

##### 3.3. Modified Bidirectional Single Shooting Method

Single shooting methods are one of the most commonly used numerical methods to solve HBVPs, which attempt to identify appropriate initial conditions for a related initial value problem to provide the solution of the original boundary value problem. If the boundary conditions are not satisfied to the desired accuracy, the selecting process of initial conditions is repeated with a new set of initial conditions until the desired accuracy is achieved. Single shooting methods are typically classified as forward single shooting method (FSSM), backward single shooting method (BSSM), and bidirectional single shooting method (BISSM), according to their different integration directions. As illustrated in Figure 1, FSSM and BSSM, which start from the initial and terminal boundary points, respectively, are unidirectional. Their solving procedures are all the same except for the opposite integral direction. Other than the above unidirectional shooting methods, BISSM starts from both boundary points and shoots to an arbitrary intermediate point [31]. However, these methods are not appropriate to solve HOCPs, mainly due to the extreme difficulty of selecting proper initial guess, even if multiple-precision arithmetic and Taylor series method are incorporated.