Hamilton Jacobi Bellman and uniqueness

16 Views Asked by At

In the framework of optimal control we have introduced in class the HJB equation to solve optimal control problem of the form

$$ \inf_{u\in\mathcal{U}}\left[\int_{0}^{T}L(t,x_u(t),u(t))dt + h(x_u(T));\quad x’(t)=f(t,x(t),u(t)), x(0)=x_0\right] $$

Where $x(t)\in\mathbb{R}^{d}$, $u(t)$ is a control, $L$ a cost function which is continuous, $h$ some function that is differentiable, $\mathcal{U}$ a set of admissible control and $f$ a function satisfying the hypotheses of the Cauchy Lipschitz theorem.

The interest of HJB has been described as follows : when the only theorem you know about sufficient condition (in the « Pontryagin setting ») of optimality fails, use the following result :

If the value function $v(t,x(t))$ is a solution $C^1$ of the HJB : $\partial_t v(t,x) + H^{*}(t,x(t),\partial_x v(t,x)) = 0$ and if $\forall (t,x_t)\in[0,T]\times\mathbb{R}^{d}$ there exists $u(t,x(t))$ minimizing the Hamiltonian with respect to the control variable. Then $x(t),u(t,x(t))$ is optimal in the minimization problem.

In practice, when it is solvable (linear quadratic case) we don’t explicitly right the value function but start with a guess function and see if it solved the HJB. Until now I am fine, however often when the guess was right, they say that is the value function because it solved the HJB equation, however I am not aware of the uniqueness of the solution and so that it is the value function.

So my question is the following : do we have uniqueness result for the HJB equation in the framework of optimal control ?

Thank you a lot !


The value function is defined $\forall (t,x_t)$ (where $t\in[0,T]$ and $x_t$ is an initial condition at time $t$) by

$$ v(t,x_t) =\inf_{u\in\mathcal{U}_t}\left[\int_{0}^{T}L(t,x_u(t),u(t))dt + h(x_u(T))\right] $$