Mixed ℋ
In this chapter, we discuss the mixed ℋ
The problem for linear systems was first considered by Bernstein and Haddad [67], where a solution for the output-feedback problem in terms of three coupled algebraic-Riccati-equations (AREs) was obtained by formulating it as an LQG problem with an ℋ
Another contribution to the linear literature was from Khargonekhar and Rotea [161] and Scherer et al. [238, 239], who considered more general multi-objective problems using convex optimization and/or linear-matrix-inequalities (LMI). And more lately, by Limebeer et al. [179] and Chen and Zhou [81] who considered a two-person nonzero-sum differential game approach with a multi-objective flavor (for the latter). This approach is very transparent and is reminiscent of the minimax approach to ℋ
Similarly, the nonlinear control problem has also been considered more recently by Lin [180]. He extended the results of Limebeer et al. [179], and derived a solution to the state-feedback problem in terms of a pair of cross-coupled Hamilton-Jacobi-Isaac’s equations. In this chapter, we discuss mainly this approach to the problem for both continuous-time and discrete-time nonlinear systems.
11.1 Continuous-Time Mixed ℋ
In this section, we discuss the mixed ℋ
‖K ο ∑a‖ℒ2 ≜ sup0≠w0∈S ‖z‖p‖w0‖S , |
(11.1) |
while the induced norm from w1 to z is the ℒ
‖K ο ∑a ‖ℒ∞ ≜ sup0≠w1∈P ‖z‖2‖w1‖2 |
(11.2) |
where
P ≜ {w(t):w∈ℒ∞, Rww(τ),Sww(jω) exist for all τ and all ω resp., ‖w‖p<∞},s ≜ {w(t):w∈ℒ∞, Rww(τ),Sww(jω) exist for all τ and all ω resp., ‖sww(jω)‖∞<∞}, ‖z|‖2p ≜ limT→∞ 12T T∫−T‖z(t)‖2 dt , s ‖w0‖2s = ‖Sw0w0(jω)‖∞ ,
and Rww(τ), Sww(jω) are the autocorrelation and power-spectral density matrices of w(t) respectively [152]. Notice also that, ‖(.)‖p and ‖(.)‖S
(11.3) |
such that the above induced-norms (11.1), (11.2) of the closed-loop system are minimized, and the closed-loop system is also locally asymptotically-stable. However, in this chapter, we do not solve the above problem, instead we solve an associated suboptimal ℋ
The plant is represented by an affine nonlinear causal state-space system defined on a manifold X ⊂ ℜn
∑a : {˙x = f(x)+g1(x)w+g2(x)u; x(t0)=x0z = h1(x)+k12(x)uy = x, |
(11.4) |
where x∈X
Furthermore, we assume without any loss of generality that the system (11.4) has a unique equilibrium-point at x = 0 such that f(0) = 0, h1(0) = h2(0) = 0, and for simplicity, we also make the following assumption.
Assumption 11.1.1 The system matrices are such that
hT1(x)k12(x) =0kT12(x)k12(x) =I.} |
(11.5) |
The problem can now be formally defined as follows.
Definition 11.1.1 (State-Feedback Mixed ℋ
(A) Finite-Horizon Problem (T < ∞): Find (if possible!) a time-varying static state-feedback control law of the form:
u=˜α2(x,t), ˜α2(t,0)=0, t∈ℜ,
such that:
(a) the closed-loop system
∑cla: {˙x = f(x)+g1w+g2(x)˜α2(x,t)z = h1(x)+k12(x)˜α2(x,t) |
(11.6) |
is stable with w = 0 and has locally finite ℒ2
(b) the output energy ‖z‖ℋ2
(B) Infinite-Horizon Problem (T → ∞) : In addition to the items (a) and (b) above, it is also required that
(c) the closed-loop system Σcla defined above with w ≡ 0 is locally asymptotically-stable about the equilibrium-point x = 0.
Such a problem can be formulated as a two-player nonzero-sum differential game (Chapter 2) with two cost functionals:
minu∈U,w∈W J1(u,w) = 12T∫t0(γ2‖w(τ)‖2−‖z(τ)‖2) dτ |
(11.7) |
minu∈U,w∈W J2(u,w) = 12T∫t0‖z(τ)‖2 dτ |
(11.8) |
for the finite-horizon problem, with T ≥ t0. Here, the first functional is associated with the ℋ
(11.9) |
(11.10) |
Furthermore, by minimizing the first objective with respect to w and substituting in the second objective which is then minimized with respect to u, the above pair of Nash-equilibrium strategies could be found. A sufficient condition for the solvability of the above differential game is provided from Theorem 2.3.1, Chapter 2, by the following pair of cross-coupled HJIEs for the finite-horizon state-feedback problem:
−Yt(x,t) = infw∈W {Yx(x,t)f(x,u⋆(x),w(x))+γ2‖w(x)‖2−‖z⋆(x)‖2}, Y(x,T)=0,−Vt(x,t) = minu∈U {Vx(x,t)f(x,u(x),w⋆(x))+‖z⋆(x)‖2}+0, V(x,T)=0,
for some negative-definite function Y : X→ℜ
In view of the above result, the following theorem gives sufficient conditions for the solvability of the finite-horizon problem.
Theorem 11.1.1 Consider the nonlinear system Σa defined by (11.4) and the finite-horizon SFBMH2HINLCP with cost functionals (11.7), (11.8). Suppose there exists a pair of negative and positive-definite C1-functions (with respect to both arguments) Y, V : N ×[0,T ] →ℜ
Then the state-feedback controls
(11.13) |
(11.14) |
solve the finite-horizon SFBMH2HINLCP for the system. Moreover, the optimal costs are given by
(11.15) |
(11.16) |
Proof: Assume there exists locally solutions Y < 0, V > 0 to the HJIEs (11.11), (11.12) in N⊂ X
Yt+Yx(f(x)+g1(x)w−g2(x)gT(x)VTx(x)) = 12‖h1(x)‖2+ γ22‖w−w⋆‖2 + 12‖u⋆‖2−12γ2‖w‖2 ⇔ ˙Y(x,t) = 12‖z|‖2−12γ2‖w‖2+γ22‖w−w⋆‖2 ⇔ ˙˜Y(x,t) ≤ 12γ2‖w‖2−12‖z‖2
for some function ˜Y = −Y >0
˜Y(x(T),T)−˜Y(x(t0,t0) ≤ T∫t012(γ2‖w‖2−‖z‖2)dt.
Therefore, the system has locally ℒ2−gain ≤ γ
˙x = f(x)−g2(x)gT2(x)VTx(x).
Differentiating ˜Y
˙˜Y = ˜Yt(x,t)+˜Yt(x,t)(f(x)−g2(x)gT2(x)VTx) ∀x∈N ≤ −12‖z‖2 ≤ 0.
Hence, the closed-loop system is Lyapunov-stable.
Next we prove item (b). Consider the cost functional J1(u,w) first. For any T ≥ t0, the following holds
J1(u,w)+Y(x(T),T)−Y(x(t0),t0) = T∫t0{12(γ2‖w‖2−‖z‖2)+˙Y(x,t)} dt = T∫t0{12(γ2‖w‖2−‖z|‖2)+Yt+Yx(f(x)+g1(x)w+g2(x)u}+ dt = T∫t0{ Yt+Yxf(x)Yxg2(x)u + γ22‖w+1γ2gT1(x)YTx‖2− γ22‖1γ2 gT1(x)YTx‖2 −12‖u‖2−12‖h1(x)‖2 } dt.
Using the HJIE (11.11), we have
J1(u,w)+Y(x(T),T)−Y(x(t0),t0) = T∫t0{γ22‖w+1γ2gT1(x)YTx‖2− 12‖u+gT2(x)VTx‖2+ ‖gT2(x)VTx‖2+Vx(x)g2(x)u+ Yxg2(x)(u−gT2(x)VTx(x))}dt,
and substituting u = u*, we have
J1(u⋆,w)+Y(x(T),T)−Y(x(t0),t0) = T∫t0 γ22‖w−w⋆‖dt ≥0.
Therefore
J1(u⋆,w⋆) ≤ J1(u⋆,w)
with
J1(u⋆,w⋆) = Y(x(t0),t0)
since Y (x(T),T ) = 0.
Similarly, considering the cost functional J2(u, w), the following holds for any T > 0
J2(u,w)+ V(x(T),T)−V(x(t0),t0) = T∫t0{12‖z‖2+˙V(x,t)} = T∫t0{12‖z‖2+Vt(x,t)+Vx(f(x)+ g1(x)w+g2(x)u)}dt = T∫t0{Vt+Vxf(x)+Vxg1 12‖u+gT2VTx‖2−‖gT2(x)VTx‖2+12‖hT1(x)‖2} dt.
Using the HJIE (11.12) in the above, we get
J2(u,w)+ V(x(T),T)−V(x(t0),t0) = T∫t0{12‖u+gT2VTx‖2+Vxg1(x)(w−1γ2gT1YTx)} dt.
Substituting now w = w⋆ we get
J2(u,w⋆)+ V(x(T),T)−V(x(t0),t0) = T∫t012‖u+gT2VTx‖2 dt ≥0,
and therefore
J2(u⋆,w⋆) ≤ J2(u,w⋆)
with
J2(u⋆,w⋆) =V(x(t0),t0).□
We can specialize the results of the above theorem to the linear system
∑l : {˙x = Ax+B1w+B2uz = C1x+D12wy = x |
(11.17) |
under the following assumption.
Assumption 11.1.2
CT1D12 =0, DT12D12 = I.
Then we have the following corollary.
Corollary 11.1.1 Consider the linear system Σl under the Assumption 11.1.2. Suppose there exist P1(t) ≤ 0 and P2(t) ≥ 0 solutions of the cross-coupled Riccati ordinary-differential-equations (ODEs):
− ˙P1(t) = ATP1+P1(t)A−[P1(t) P2(t)][γ−2B1BT1B2BT2B2BT2B2BT2][P1(t)P2(t)]_CT1C1, P1(T) =0− ˙P2(t) = ATP2+P2(t)A−[P1(t) P2(t)][0γ−2B1BT1γ−2B1BT1B2BT2][P1(t)P2(t)]_CT1C1, P2(T) =0
on [0,T]. Then, the Nash-equilibrium strategies uniquely specified by
u⋆ = −BT2P2(t)x(t) w⋆ = − 1γ2BT1P1(t)x(t)
solve the finite-horizon SFBMH2HICP for the system. Moreover, the optimal costs for the game associated with the system are given by
J1(u⋆,w⋆) = 12xT(t0)P1(t0)x(t0), J2(u⋆,w⋆) = 12xT(t0)P2(t0)x(t0).
Proof: Take
Y(x(t),t)=12xT(t)P1(t)x(t), P1(t) ≤0V(x(t),t)=12xT(t)P2(t)x(t), P2(t) ≥0
and apply the results of the Theorem. □
Remark 11.1.1 In the above corollary, we considered negative and positive-(semi)definite solutions of the Riccati (ODEs), while in Theorem 11.1.1 we considered strict definite solutions of the HJIEs. However, it is generally sufficient to consider semidefinite solutions of the HJIEs.
11.1.1 The Infinite-Horizon Problem
In this subsection, we consider the infinite-horizon SFBMH2HINLCP for the affine nonlinear system Σa. In this case, we let T → ∞, and seek time-invariant functions and feedback gains that solve the HJIEs. Because of this, it is necessary to require that the closed-loop system is locally asymptotically-stable as stated in item (c) of the definition. However, to achieve this, some additional assumptions on the system might be necessary. The following theorem gives sufficient conditions for the solvability of this problem. We recall the definition of detectability first.
Definition 11.1.2 The pair {f,h} is said to be locally zero-state detectable if there exists a neighborhood O
Theorem 11.1.2 Consider the nonlinear system Σa defined by (11.4) and the infinitehorizon SFBMH2HINLCP with cost functions (11.7), (11.8). Suppose
(H1) the pair {f,h1} is zero-state detectable;
(H2) there exists a pair of negative and positive-definite C1-functions ˜Y,˜V : ˜N → ℜ
˜Vx(x)f(x)−12˜Vx(x)g2(x)gT2(x)˜VTx(x)−1γ2˜Vx(x)g1(x)gT1(x)˜YTx(x)+ 12hT1(x)h1(x)=0, ˜V(0)=0. |
(11.19) |
Then the state-feedback controls
(11.20) |
(11.21) |
solve the infinite-horizon SFBMH2HINLCP for the system. Moreover, the optimal costs are given by
(11.22) |
(11.23) |
Proof: We only prove item (c) in the definition, since the proofs of items (a) and (b) are similar to the finite-horizon case. Using similar manipulations as in the proof of item (a) of Theorem 11.1.1, it can be shown that with w ≡0,
˙⌣Y = −12‖z‖2,
for some function ⌣Y = −˜Y
The above theorem can again be specialized to the linear system Σl in the following corollary.
Corollary 11.1.2 Consider the linear system Σl under the Assumption 11.1.2. Suppose (C1, A) is detectable and there exists symmetric solutions ˉP1 ≤ 0 and ˉP2 ≥ 0
ATˉP1+ˉP1A−[ˉP1 ˉP2][γ−2B1BT1B2BT2B2BT2B2BT2][ˉP1 ˉP2+]_CT1C1=0,
ATˉP2+ˉP2A−[ˉP1 ˉP2][0γ−2B1BT1γ−2B1BT1B2BT2][ˉP1 ˉP2]+CT1C1=0
Then, the Nash-equilibrium strategies uniquely specified by
(11.24) |
(11.25) |
solve the infinite-horizon SFBMH2HINLCP for the system. Moreover, the optimal costs for the game associated with the system are given by
J1(u⋆,w⋆) = 12xT(t0)ˉP1x(t0), |
(11.26) |
J2(u⋆,w⋆) = 12xT(t0)ˉP2x(t0). |
(11.27) |
Proof: Take
Y(x,t) = 12xTˉP1x, ˉP1≤0, V(x,t)=12xTˉP2x, ˉP2≤0.
and apply the result of the theorem. □
Remark 11.1.2 Sufficient conditions for the existence of the asymptotic solutions to the coupled algebraic-Riccati equations are discussed in Reference [222].
The following proposition can also be proven.
Proposition 11.1.1 Consider the nonlinear system (11.4) and the infinite-horizon SFBMH2HINLCP for this system. Suppose the following hold:
(a1) {f + g1w⋆, h1} is locally zero-state detectable;
(a2) there exists a pair of C1 negative and positive-definite functions ˜Y ,˜V
Then f−g2gT2˜VTx−1γ2g1gT1˜YTx
Proof: Rewrite the HJIE (11.19) as
(f(x)−g2(x)gT2(x)˜VTx(x)−12gT1(x)gT1(x)˜Yx(x)) = −12‖gT2(x)˜VTx(x)‖2−12‖h1(x)‖2 ⇔ ˙˜V = −12‖gT2(x)˜VTx(x)‖2−12‖h1(x)‖2 ≤ 0 .
Again the condition ˙˜V ≡ 0 ∀ t≥ts
Remark 11.1.3 The proof of the equivalent result for the linear system Σl can be pursued along the same lines. Moreover, many interesting corollaries could also be derived (see also Reference [179]).
11.1.2 Extension to a General Class of Nonlinear Systems
In this subsection, we consider the SFBMH2HINLCP for a general class of nonlinear systems which are not necessarily affine, and extend the approach developed above to this class of systems. We consider the class of nonlinear systems described by
∑ : {˙x = F(x,w,u), x(t0)=x0z = Z(x,u)y = x |
(11.28) |
where all the variables have their previous meanings and dimensions, while F: X×W×U→ℜn , Z: X×U→ℜs
Assumption 11.1.3 The system matrices satisfy the following conditions:
(i) det [(∂Z∂u(0,0))T (∂Z∂u(0,0))] ≠ 0;
(ii) Z(x,u) = 0 ⇒ u=0
Define now the Hamiltonian functions Ki: T⋆ X×W×U→ℜ, i=1,2
K1(x,w,u,ˉYTx) = ˉYx(x) F(x,w,u) +12γ2‖w‖2−‖z(x,u)‖2 K2(x,w,u,ˉVTx) = ˉVx(x) F(x,w,u) +12‖z(x,u)‖2
for some smooth functions ˉY,ˉV : X→ℜ, and where p1= ˉYTx, p2 = ˉVTx
∂2K|w=0,u=0(0) ≜ [∂2K2∂u2∂2K2∂w∂u∂2K1∂u∂w∂2K1∂w2]|w=0,u=0 (0) = [(∂Z∂u(0,0))T (∂Z∂u(0,0))00γ2I]
is nonsingular by Assumption 11.1.3. Therefore, by the Implicit-function Theorem, there exists an open neighborhood ˉX
∂K1∂w(x,ˉw⋆(x),ˉu⋆(x)) = 0,
∂K2∂u(x,ˉw⋆(x),ˉu⋆(x)) = 0
have unique solutions (ˉu⋆(x),ˉw⋆(x)), with ˉu⋆(0)=0,ˉw⋆(0)=0
Theorem 11.1.3 Consider the nonlinear system (11.28) and the SFBMH2HINLCP for it. Suppose Assumption 11.1.3 holds, and also the following:
(A1) the pair {F (x, 0, 0),Z(x,0)} is zero-state detectable;
(A2) there exists a pair of C1 locally negative and positive-definite functions ˉY,ˉV : ˉN→ℜ
respectively, defined in a neighborhood ˉN
ˉYx(x)F(x,ˉw⋆(x),ˉu⋆(x))+12γ2‖ˉw⋆(x)‖2−12‖Z(x,ˉu⋆(x))‖2 = 0, ˉY(0) =0,
ˉVx(x)F(x,ˉw⋆(x),ˉu⋆(x))+12‖Z(x,ˉu⋆(x))‖2 = 0, ˉV(0)=0;
(A3) the pair {F(x,ˉw⋆(x),0),Z(x,0)}
Then, the state-feedback controls (ˉu⋆(x),ˉw⋆(x))
ˉJ⋆1(ˉu⋆,ˉw⋆) = ˉY(x0), ˉJ⋆2(ˉu⋆,ˉw⋆) = ˉV(x0).
Proof: The proof can be pursued along the same lines as the previous results. □