最优控制原理

性能泛函
$$J = \int_{t_0}^{t_f} F(x,\dot{x},t) dt.$$
其中，$t_0$，$x(t_0)$固定.

若$t_f$固定，则
\begin{equation}
\nonumber
\begin{split}
\delta_x J = & \int_{t_0}^{t_f} \left( \frac{\partial F}{\partial x} \delta x + \frac{\partial F}{\partial \dot{x}} \delta \dot{x} \right) dt \\
= & \int_{t_0}^{t_f} \left( \frac{\partial F}{\partial x}  - \frac{d}{dt} \frac{\partial F}{\partial \dot{x}} \right)\delta x dt + \frac{\partial F}{\partial \dot{x}} \delta x \bigg\vert_{t_f}.
\end{split}
\end{equation}
显然，上式等于零，则
$$\frac{\partial F}{\partial x}  - \frac{d}{dt} \frac{\partial F}{\partial \dot{x}} = 0, \quad t_0 \leq t \leq t_f.$$
$$\frac{\partial F}{\partial \dot{x}} \bigg\vert_{t_f} \delta x(t_f) = 0.$$
所以，若$x(t_f)$固定，则第二约束必然成立.若$x(t_f)$自由，则
$$\frac{\partial F}{\partial \dot{x}} \bigg\vert_{t_f} = 0.$$

若$t_f$自由，终端满足约束
$$x(t_f) = c(t_f).$$
即函数$c(t)$表示，以时刻$t$为终端时刻的情况下，终端状态应当等于$c(t)$.此时
$$J_a = \int_{t_0}^{t_f} F(x,\dot{x},t) dt + \lambda(c(t_f) - x(t_f)).$$
所以
$$\delta_t J_a = F(t_f) \delta t + \lambda (\dot{c}(t_f) - \dot{x}(t_f)) \delta t.$$
$$\delta_x J_a = \int_{t_0}^{t_f} \left( \frac{\partial F}{\partial x}  - \frac{d}{dt} \frac{\partial F}{\partial \dot{x}} \right)\delta x dt + \frac{\partial F}{\partial \dot{x}} \bigg\vert_{t_f} \delta x(t_f) - \lambda \delta x(t_f).$$
$$\delta_{\lambda} J_a = (c(t_f) - x(t_f)) \delta \lambda.$$
显然，上式等于零，则
$$F(t_f) + \lambda (\dot{c}(t_f) - \dot{x}(t_f)) = 0.$$
$$\frac{\partial F}{\partial x}  - \frac{d}{dt} \frac{\partial F}{\partial \dot{x}} = 0, \quad t_0 \leq t \leq t_f.$$
$$\frac{\partial F}{\partial \dot{x}} \bigg\vert_{t_f} - \lambda = 0.$$
$$c(t_f) - x(t_f) = 0.$$

系统方程
$$\dot{x} = f(x, u, t).$$
性能泛函
$$J = \phi(x(t_f), t_f) + \int_{t_0}^{t_f} F(x,u,t) dt.$$
其中，$t_0$，$x(t_0)$固定.

令
$$H(x,u,\lambda,t) = F(x,u,t) + \lambda f(x,u,t).$$
若$t_f$固定，则
\begin{equation}
\nonumber
\begin{split}
J_a = & \phi(x(t_f), t_f) + \int_{t_0}^{t_f} (F(x,u,t) + \lambda(f(x,u,t) - \dot{x})) dt \\
= & \phi(x(t_f), t_f) + \int_{t_0}^{t_f} (H(x,u,\lambda,t) - \lambda \dot{x}) dt.
\end{split}
\end{equation}
所以
\begin{equation}
\nonumber
\begin{split}
\delta_x J_a = & \frac{\partial \phi}{\partial x} \bigg\vert_{t_f} \delta x(t_f) + \int_{t_0}^{t_f} \left( \frac{\partial H}{\partial x} \delta x - \lambda \delta \dot{x} \right) dt \\
= & \frac{\partial \phi}{\partial x} \bigg\vert_{t_f} \delta x(t_f) + \int_{t_0}^{t_f} \left( \frac{\partial H}{\partial x}  + \dot{\lambda} \right)\delta x dt - \lambda(t_f) \delta x(t_f).
\end{split}
\end{equation}
$$\delta_{u} J_a = \int_{t_0}^{t_f} \frac{\partial H}{\partial u}  \delta u dt.$$
$$\delta_{\lambda} J_a = \int_{t_0}^{t_f} \left( \frac{\partial H}{\partial \lambda}  - \dot{x} \right) \delta \lambda dt.$$
显然，上式等于零，则
$$\frac{\partial H}{\partial x}  + \dot{\lambda} = 0, \quad t_0 \leq t \leq t_f.$$
$$\left( \frac{\partial \phi}{\partial x} \bigg\vert_{t_f} - \lambda(t_f) \right) \delta x(t_f) = 0.$$
$$\frac{\partial H}{\partial u} = 0, \quad t_0 \leq t \leq t_f.$$
$$\frac{\partial H}{\partial \lambda}  - \dot{x} = 0, \quad t_0 \leq t \leq t_f.$$
所以，若$x(t_f)$固定，则第二约束必然成立.若$x(t_f)$自由，则
$$\frac{\partial \phi}{\partial x} \bigg\vert_{t_f} - \lambda(t_f) = 0.$$

若$t_f$自由，终端满足约束
$$G(x(t_f), t_f) = 0.$$
令
$$\theta(x(t_f), t_f, v) = \phi(x(t_f), t_f) + v G(x(t_f), t_f).$$
则
$$J_a = \theta(x(t_f), t_f, v) + \int_{t_0}^{t_f} (H(x,u,\lambda,t) - \lambda \dot{x}) dt.$$
所以
$$\delta_x J_a = \frac{\partial \theta}{\partial x} \bigg\vert_{t_f} \delta x(t_f) + \int_{t_0}^{t_f} \left( \frac{\partial H}{\partial x}  + \dot{\lambda} \right)\delta x dt - \lambda(t_f) \delta x(t_f).$$
$$\delta_{u} J_a = \int_{t_0}^{t_f} \frac{\partial H}{\partial u}  \delta u dt.$$
\begin{equation}
\nonumber
\begin{split}
\delta_t J_a = & \frac{\partial \phi}{\partial x} \bigg\vert_{t_f} \dot{x}(t_f) \delta t + \frac{\partial \phi}{\partial t} \bigg\vert_{t_f} \delta t + (H(t_f) - \lambda(t_f) \dot{x}(t_f)) \delta t \\
= & \frac{\partial \phi}{\partial t} \bigg\vert_{t_f} \delta t + H(t_f) \delta t.
\end{split}
\end{equation}
$$\delta_{\lambda} J_a = \int_{t_0}^{t_f} \left( \frac{\partial H}{\partial \lambda}  - \dot{x} \right) \delta \lambda dt.$$
$$\delta_{v} J_a = \frac{\partial \theta}{\partial v} \bigg\vert_{t_f} \delta v.$$
显然，上式等于零，则
$$\frac{\partial H}{\partial x}  + \dot{\lambda} = 0, \quad t_0 \leq t \leq t_f.$$
$$\frac{\partial \phi}{\partial x} \bigg\vert_{t_f} - \lambda(t_f) = 0.$$
$$\frac{\partial H}{\partial u} = 0, \quad t_0 \leq t \leq t_f.$$
$$\frac{\partial \phi}{\partial t} \bigg\vert_{t_f} \delta t + H(t_f) = 0.$$
$$\frac{\partial H}{\partial \lambda}  - \dot{x} = 0, \quad t_0 \leq t \leq t_f.$$
$$\frac{\partial \theta}{\partial v} \bigg\vert_{t_f} = 0.$$

对于$u$受约束时
$$u \in \Omega.$$
显然
$$\min_{u \in \Omega} J_a.$$
其余变量的微分条件不变.

令
$$V(x, t) = \min_{u \in \Omega} \left\{ \phi(x(t_f), t_f) + \int_{t}^{t_f} F(x,u,t) dt \right\}.$$
其中，$V(x,t)$的$x$表示时刻$t$的状态.所以
\begin{equation}
\nonumber
\begin{split}
V(x,t) = & \min_{u \in \Omega} \left\{ \int_{t}^{t + \delta t} F dt + V(x + \delta x, t + \delta t) \right\} \\
= & \min_{u \in \Omega} \left\{ F(t) \delta t + V(x, t) + \frac{\partial V}{\partial x} f(t) \delta t + \frac{\partial V}{\partial t} \delta t \right\} \\ 
= & \min_{u \in \Omega} \left\{ F(t) \delta t + \frac{\partial V}{\partial x} f(t) \delta t \right\} + V(x, t) + \frac{\partial V}{\partial t} \delta t.
\end{split}
\end{equation}
所以
$$- \frac{\partial V}{\partial t} = \min_{u \in \Omega} \left\{ F(t) + \frac{\partial V}{\partial x} f(t) \right\} = \min_{u \in \Omega} H\left(x, u, \frac{\partial V}{\partial x}, t\right).$$
边界条件为
$$V(x(t_f), t_f) = \phi(x(t_f), t_f).$$
所以
$$\lambda = \frac{\partial V}{\partial x}.$$
称为最优原理.

线性二次型
$$\dot{x} = Ax + Bu.$$
$$J = \frac{1}{2} x^T(t_f) P(t_f) x(t_f) + \frac{1}{2} \int_{t_0}^{t_f} [x^T(t)Q(t)x(t) + u^T(t)R(t)u(t)] dt.$$
所以
$$H= \frac{1}{2}x^TQx + \frac{1}{2}u^TRu + \lambda^T(Ax+Bu).$$
$$\theta = \frac{1}{2} x^TPx.$$
假设$t_0$，$t_f$固定，则
$$\frac{\partial H}{\partial x} + \dot{\lambda} = Qx + A^T \lambda + \dot{\lambda} = 0.$$
$$\frac{\partial H}{\partial u} = Ru + B^T\lambda = 0.$$
$$\frac{\partial H}{\partial \lambda} - \dot{x} = Ax+Bu - \dot{x} = 0.$$
$$\frac{\partial \theta}{\partial x} \bigg\vert_{t_f} - \lambda(t_f) = P(t_f) x(t_f) - \lambda(t_f) = 0.$$
假设
$$\lambda = Kx.$$
则
$$\dot{K} = KBR^{-1}B^TK -KA -A^TK - Q.$$
$$K(t_f) = P(t_f).$$
称为Riccati方程.显然
$$K^T = K.$$
$$u = -R^{-1}B^TKx.$$

若系统为线性定常系统，并且趋于稳定，$t_f = \infty$，则显然
$$\dot{K} = 0.$$
即
$$KBR^{-1}B^TK -KA -A^TK - Q = 0.$$

离散系统
$$x(k+1) = f(x(k), u(k), k), \quad x(0) = x_0.$$
终端约束
$$G(x(N), N) = 0.$$
性能指标
$$J = \phi(x(N), N) + \sum_{k=0}^{N-1} F(x(k), u(k). k).$$
则
$$H(k) = F(x(k), u(k). k) + \lambda(k+1) f(x(k), u(k), k).$$
$$\theta(N) = \phi(x(N), N) + vG(x(N), N).$$
所以
$$J_a = \theta(N) + \sum_{k=0}^{N-1} H(k) - \lambda(k+1)x(k+1).$$
所以
$$\delta_{v} J_a = \frac{\partial \theta}{\partial v} \delta v = 0.$$
$$\delta_{\lambda} J_a = \sum_{k=0}^{N-1} \frac{\partial H(k)}{\partial \lambda} \delta \lambda(k+1) - x(k+1) \delta \lambda(k+1) = 0.$$
$$\delta_{x} J_a = \frac{\partial \theta}{\partial x} \delta x(N) + \sum_{k=0}^{N-1} \frac{\partial H(k)}{\partial x} \delta x(k) - \lambda(k+1) \delta x(k+1) = 0.$$
$$\delta_{u} J_a = \sum_{k=0}^{N-1} \frac{\partial H(k)}{\partial u} \delta u(k) = 0.$$
所以
$$\frac{\partial \theta}{\partial v} = 0.$$
$$\frac{\partial H(k)}{\partial \lambda} - x(k+1) = 0, \quad k=0, 1, \ldots, N-1.$$
$$\frac{\partial H(k)}{\partial x} - \lambda(k) = 0, \quad k=1, 1, \ldots, N-1.$$
$$\frac{\partial \theta}{\partial v} \bigg\vert_{N} - \lambda(N) = 0.$$
$$\frac{\partial H(k)}{\partial u} = 0, k=0,1,\ldots,N-1.$$
显然，对于$N$的变化是离散的.

离散线性二次型
$$x(k+1) = A(k) x(k) + B(k) u(k).$$
$$J= \frac{1}{2} x^T(N) P(N) x(N) + \frac{1}{2} \sum_{k=0}^{N-1} [x^T(k)Q(k)x(k) + u^T(k)R(k)u(k)].$$
假设
$$\lambda(k) = K(k) x(k).$$
则
$$K(k) = Q(k) + A^T(k)K(k+1)[I + B(k)R^{-1}(k)B^T(k)K(k+1)]^{-1}A(k).$$
$$K(N) = P(N).$$
所以
$$u(k) = -[R(k) + B^T(k)K(k+1)B(k)]^{-1} B^T(k)K(k+1)A(k) x(k).$$