B4.14

Optimization and Control | Part II, 2001

Consider the scalar system with plant equation xt+1=xt+ut,t=0,1,x_{t+1}=x_{t}+u_{t}, t=0,1, \ldots and cost

Cs(x0,u0,u1,)=t=0s[ut2+43xt2]C_{s}\left(x_{0}, u_{0}, u_{1}, \ldots\right)=\sum_{t=0}^{s}\left[u_{t}^{2}+\frac{4}{3} x_{t}^{2}\right]

Show from first principles that minu0,u1,Cs=Vsx02\min _{u_{0}, u_{1}, \ldots} C_{s}=V_{s} x_{0}^{2}, where V0=4/3V_{0}=4 / 3 and for s=0,1,s=0,1, \ldots

Vs+1=4/3+Vs/(1+Vs)V_{s+1}=4 / 3+V_{s} /\left(1+V_{s}\right)

Show that Vs2V_{s} \rightarrow 2 as ss \rightarrow \infty.

Prove that CC_{\infty} is minimized by the stationary control, ut=2xt/3u_{t}=-2 x_{t} / 3 for all tt.

Consider the stationary policy π0\pi_{0} that has ut=xtu_{t}=-x_{t} for all tt. What is the value of CC_{\infty} under this policy?

Consider the following algorithm, in which steps 1 and 2 are repeated as many times as desired.

  1. For a given stationary policy πn\pi_{n}, for which ut=knxtu_{t}=k_{n} x_{t} for all tt, determine the value of CC_{\infty} under this policy as Vπnx02V^{\pi_{n}} x_{0}^{2} by solving for VπnV^{\pi_{n}} in

Vπn=kn2+4/3+(1+kn)2VπnV^{\pi_{n}}=k_{n}^{2}+4 / 3+\left(1+k_{n}\right)^{2} V^{\pi_{n}}

  1. Now find kn+1k_{n+1} as the minimizer of

kn+12+4/3+(1+kn+1)2Vπnk_{n+1}^{2}+4 / 3+\left(1+k_{n+1}\right)^{2} V^{\pi_{n}}

and define πn+1\pi_{n+1} as the policy for which ut=kn+1xtu_{t}=k_{n+1} x_{t} for all tt.

Explain why πn+1\pi_{n+1} is guaranteed to be a better policy than πn\pi_{n}.

Let π0\pi_{0} be the stationary policy with ut=xtu_{t}=-x_{t}. Determine π1\pi_{1} and verify that it minimizes CC_{\infty} to within 0.2%0.2 \% of its optimum.

Typos? Please submit corrections to this page on GitHub.