庞特里亚金原理 (Pontryagin‘s Principle)

yyakow

已于 2024-04-10 10:44:46 修改

阅读量1.1k

点赞数 6

分类专栏：最优控制的学习笔记文章标签：算法

于 2024-04-07 15:00:32 首次发布

本文链接：https://blog.csdn.net/yyakow/article/details/137462691

版权

最优控制的学习笔记专栏收录该内容

10 篇文章 0 订阅

订阅专栏

话不多说，先给出文献的英文版定理，可能自己理解之后会重新整理一下该定理。

Theorem. (Pontryagin's Principle) Necessary conditions that $(x_{0}^{*}, u^{*}(t))$ be an optimal initial condition and optimal control for the optimal control problem are the existence of a nonzero k-dimensional vector $\lambda$ with $\lambda_{1}\leq 0$ and an n-dimensional vector function $P(t)$ such that for $t \in [t_{0}, t_{1}]$ :

$\dot{P(t)}^{\prime} = -P(t)^{\prime} f_{x}(t, x^{*}(t), u^{*}(t))$ ; (1)

for $t \in[t_{0}, t_{1}]$ and $u\in U$

$P(t)^{\prime}[f(t, x^{*}(t), u)-f(t, x^{*}(t), u^{*}(t))]\leq 0$ ; (2)

$P(t_{1})^{\prime}= \lambda^{\prime} \phi_{x_{1}}(e)$ ; (3)

$P(t_{0})^{\prime}= -\lambda^{\prime} \phi_{x_{0}}(e)$ ; (4)

$P(t_{1})^{\prime} f(t_{1}, x^{*}(t_{1}), u^{*}(t_{1}))= -\lambda^{\prime} \phi_{t_{1}}(e)$ ; (5)

$P(t_{0})^{\prime} f(t_{0}, x^{*}(t_{0}), u^{*}(t_{0}))= \lambda^{\prime} \phi_{t_{0}}(e)$ . (6)

If $f(t, x, u)$ has a continuous partial derivation $f_{t}(t, x, u)$ , then the condition

$P(t)^{\prime} f(t, x^{*}(t), u^{*}(t))=\lambda^{\prime} \phi_{t_{0}}(t_{0}, t_{1}, x^{*}(t_{0}), x^{*}(t_{1})) + \int\limits_{t_{0}}^{t_{1}} P(s)^{\prime} f_{t}(S, x^{*}(s), u^{*}(s))d s$ (7)

holds for each $t \in [t_{0}, t_{1}]$ .

The quantity

$H(t, x, u)=P(t)^{\prime} f(t, x, u)$ (8)

is generally called the Hamiltonian in analogy with a corresponding quantity occuring in classic mechanics. Condition (2) can be expressed as

$\max\limits_{u\in U}\{ H(t, x^{*}(t), u)\}=H(t, x^{*}(t), u^{*}(t))$ (9)

and is called Pontryagin's maximum principle. Conditions (3)-(6) called transversality conditions (横截条件). Eq. (1) are called the adjoint equations (伴随方程).

由于庞特里亚金原理的条件是最优性的必要条件, 因此每个最优控制都必须是极值. 但是, 由于条件不一定是最优的, 因此可能存在非最优的极端控制。

问题是, 为什么我所阅读的本专业的文献最优控制就是利用庞特里亚金原理得到的呢? 它明明只是必要条件啊?? 是我忽略了什么信息吗?? 还是说这与利用庞特里亚金原理直接设置数值计算方法求解最优控制 $u^{*}$ 完全不矛盾??

功夫不负有心人, 上面的疑惑得到了一部分的解答. 虽然 Pontryagin's Principle 只是最优控制的必要条件, 但是当对结束状态不受限制时 (the free terminal point problem), 在一些特殊情形下, Pontryagin's Principle 就会变成最优控制的充分必要条件, 详细的说明如下:

Let $U$ be convex. Let the equations of motion of the system be given by the linear differential equation

$\dot{x}=A(t)x(t)+B(t)u(t)$ (10)

where $A(t)$ and $B(t)$ are appropriate dimensional matrices of continuous functions. Let the performance index be given by

$J(u)=\int\limits_{t_{0}}^{t_{1}} L(t, x(t), u(t)) dt+\psi(x(t_{1}))$ (11)

where $L$ is a continuous real valued function continuously differentiable and convex in $(x, u)$ and $\psi$ is a continously differentiable convex function of $x$ . Under these assumptions the conditions can be rewritten:

Theorem. A necessary and sufficient condition for optimality of a control $u(t)$ for the free terminal point problem with system Eq. (10) and performance index (11) is that for $t \in (t_{0}, t_{1}]$

$-L_{u}(t, x(t), u(t)) v+\tilde{P}(t)^{\prime} B(t) v \leq 0$ (12)

for each $v \in U$ such that $u(t)+v\in U$ , where $\tilde{P}(t)$ is the solution of

$\dot{\tilde{P}}(t)^{\prime}=-\tilde{P}(t)^{\prime} A(t)+L_{x}(t, x(t), u(t))$ (13)

$\tilde{P}(t_{1})^{\prime}=-\psi_{x}(x(t_{1}))$ . (14)

Moreover if in addition $L(t, x, u)$ is strictly convex in $(x, u)$ for each fixed $t$ , the optimal control $u(t)$ is unique.

所以, 如果控制集 $U$ 是凸的, 并且 $L$ 也是凸的, 那么 Pontryagin's Principle 就是充分必要条件. 不过这个定理是针对线性微分方程的, 非线性微分方程是否也具有一样的性质? 如果是的话, 我所阅读的文献的确满足这些假设条件, 因此, 利用 Ponytryagin's Principle 来计算最优控制也是毋庸置疑的.