=Paper=
{{Paper
|id=Vol-1987/paper9
|storemode=property
|title=About One Control Problem with Incomplete Information
|pdfUrl=https://ceur-ws.org/Vol-1987/paper9.pdf
|volume=Vol-1987
|authors=Boris I. Ananyev
}}
==About One Control Problem with Incomplete Information==
https://ceur-ws.org/Vol-1987/paper9.pdf
About One Control Problem with Incomplete
Information
Boris I. Ananyev
N.N. Krasovskii Institute of Mathematics and Mechanics UB of RAS
Kovalevskaya str. 16,
620990 Yekaterinburg, Russia.
abi@imm.uran.ru
Abstract
A linear control system describing the deviation from a rated trajec-
tory is considered. The state vector is unobserved and the controller
can watch only an output with disturbances constrained in LQ-norm.
Using the theory of guaranteed estimation, the controller (1st player)
builds informational sets and tries to minimize the payoff functional
depending of the informational set at the end of observation process.
The second player (an opponent) choosing the disturbances intends to
prevent the controller and tries to maximize the functional. The situa-
tion is reduced to a differential game with complete information where
the players are used the positional strategies depending on parameters
of informational sets. An example with a flying vehicle is examined.
1 Introduction
We use the theory of guaranteed estimation [Kurzhanski & Varaiya, 2014] and give a generalization for the case
with integral constraints. As the result of the solution we obtain the differential equations describing the evolution
of the informational sets. With the help of these equations the differential game with complete information is
formulated. To solve this we use the theory of positional differential games [Krasovskii, 1985, Lokshin, 1992,
Souganidis, 1985, Souganidis, 1999, Subbotin, 1999]. The existence of an optimal strategy is established and a
method of construction is suggested. It is established that the differential game has a value and a saddle point.
Another approach to the problem under consideration is suggested in [Ananyev, 2017]. Observations’ control
problems were considered in [Ananiev, 2011, Ananyev, 2012].
The paper is organized as follows. In section 2 we give the background for guaranteed estimation. Section
3 contains a formulation of the problem. In section 4 we consider a solution with the help of Hamilton-Jacobi-
Bellman-Isaacs (HJBI) equation and suggest a numerical scheme. Here we define the strategies of players and
stepwise solutions of the equations according to [Krasovskii, 1985, Subbotin, 1999]. Unfortunately, we cannot
solve the HJBI equation with unbounded functions of the 2nd player. But we avoid this temporarily introducing
the constraint on functions of the 2nd player. After that we pass to the limit and obtain the value of the game
which computed in [Lokshin, 1992] by other way. Section 5 is devoted to an example.
Copyright ⃝
c by the paper’s authors. Copying permitted for private and academic purposes.
In: Yu. G. Evtushenko, M. Yu. Khachay, O. V. Khamisov, Yu. A. Kochetov, V.U. Malkova, M.A. Posypkin (eds.): Proceedings of
the OPTIMA-2017 Conference, Petrovac, Montenegro, 02-Oct-2017, published at http://ceur-ws.org
52
2 Guaranteed Estimation
Consider linear non-stationary equations
ẋ(t) = A(t)x(t) + B(t)u(t) + C(t)v(t), (1)
y(t) = G(t)x(t) + c(t)v(t), t ∈ [0, T ], (2)
containing an uncertain function v(·) and a control u(·), where A(·), B(·), C(·), G(·), c(·) are bounded Borelean
matrices. The unobserved state vector x(t) ∈ Rn , the output y(t) ∈ Rm , u(t) ∈ Rp , v(t) ∈ Rk . Suppose that the
uncertain function v(·) in (1) and (2) is constrained by the inequality
∫ T
∥v(·)∥2 = |v(t)|2 dt ≤ 1, (3)
0
where | · | is the Euclidean norm. Besides, the matrix c(·) must satisfy the condition
c(t)c′ (t) ≥ δIm , ∀t ∈ [0, T ], (4)
where δ > 0 and Im ∈ Rm×m is the identity matrix. Hereafter the symbol ′ means the transposition. According
to general theory of guaranteed estimation [Kurzhanski & Varaiya, 2014], we give
Definition 1 The collection XT (u, y) of state vectors {x(T )} is said to be the informational set if for any
x ∈ XT (u, y) there exists a function v(·) satisfying (3) such that equality (2) holds a.e.
Denote by C(t) the matrix (c(t)c′ (t))−1 . Under assumption (4) we have the equalities v(t) = c′ (t)w(t) + C1 (t)f (t)
and ∥v(·)∥2 = ∥c′ (·)w(·)∥2 + ∥C1 (·)f (·)∥2 , where C1 (t) = Ik − c′ (t)C(t)c(t) is the orthogonal projection on the
subspace ker c(t). Using (2), we may rewrite inequality (3) as
∫ T{ }
|y(t) − G(t)x(t)|2C(t) + |f (t)|2C1 (t) dt ≤ 1. (5)
0
From now on, the symbol |x|2Q means x′ Qx, where Q is a symmetrical and non-negatively defined matrix.
It is easily seen that x ∈ XT (u, y) iff there exists a function f (·) satisfying (5) and subjecting to the equation
ẋ(t) = A(t)x(t) + B(t)u(t) + C(t)(C1 (t)f (t) + c′ (t)C(t)(y(t) − G(t)x(t)))
with final condition x(T ) = x. On the other hand, such a function exists iff the minimum on f (·) of the left-hand
side of inequality (5) is less or equal 1. To find this minimum we use the Bellman equation for an equation
ẋ = f (t, x, v), v(t) ∈ P, t ∈ [0, T ], x(T ) = xT ,
with a cost functional ∫ T
J(T, xT ) = L(t, x, v)dt + ϕ(x0 ).
0
For any end position {t, x} we find
{∫ t }
J(t, x) = inf L(s, x(s), v(s))ds + ϕ(x0 ) , x(t) = x.
v(·) 0
Bellman’s equation in partial derivatives is of the form:
{ }
Jt = min L(t, x, v) − Jx f (t, x, v) , J(0, x) = ϕ(x).
v∈P
Let the solution J(t, x) of Bellman’s equation be continuous in t, x and differentiable in x. Then it gives
the minimum of the functional. This fact is sufficient for us. Note that there is a more advanced theory of
viscosity solutions, [Fleming & Soner, 2006]. Trying to solve the Bellman’s equation with a quadratic form
|x|2P (t) − 2d′ (t)x + q(t), we come to the conclusion.
53
{ }
Lemma 1 The informational set has the form XT (u, y) = x ∈ Rn : |x|2P (T ) − 2d′ (T )x + q(T ) ≤ 1 , where
parameters may be found from equations
Ṗ (t) = −P (t)A(t) − A′ (t)P (t) − P (t)C(t)C ′ (t)P (t) + (G(t) + c(t)C ′ (t)P (t))′ C(t)(G(t) + c(t)C ′ (t)P (t)),
(6)
P (0) = 0;
( )′
˙ = − A(t) + C(t)C (t)P (t) d(t) + (G(t) + c(t)C ′ (t)P (t))′ C(t)(y(t) + c(t)C ′ (t)d(t)) + P (t)B(t)u(t),
d(t) ′
(7)
d(0) = 0;
2 2
q̇(t) = y(t) C(t) − C ′ (t)d(t)| C1 (t) + 2d′ (t)B(t)u(t), q(0) = 0. (8)
If the matrix P (t) is invertible on (0, T ], we can introduce the values x̂(t) = P −1 (t)d(t) and h(t) = q(t) −
2
d(t) P −1 (t) which satisfy the equations
(
˙
x̂(t) = A(t)x̂(t) + c(t)C ′ (t) + G(t)P −1 (t))′ C(t)(y(t) − G(t)x̂(t)) + B(t)u(t), (9)
2
ḣ(t) = y(t) − G(t)x̂(t) C(t) . (10)
The value x̂(T ) is the center of bounded ellipsoid XT (u, y). A simple sufficient condition for invertibility of
matrix P (t) on (0, T ] is the following.
Assumption 1 The solutions W (t, τ ) of matrix differential equation
∂W (t, τ )/∂t = A′ (t)W (t, τ ) + W (t, τ )A(t) + G′ (t)G(t), W (τ, τ ) = 0,
is positive-definite for all 0 ≤ τ < t ≤ T .
It is well-known (see [Kurzhanski & Varaiya, 2014]) that Assumption 1 is equivalent to full observability of
system (1), (2), where u = 0, v = 0, on any interval [τ, t].
Remark
The calculation of value h(t) is extremely unstable. Practically, one can take a small matrix P0 > 0 and solve
equations (9), (10) with zero initial data. If h(t) from (10) is grater 1, one needs to take a smaller matrix.
Another way to overcome this difficulty is to set ν0 = 0. This increases a bit the initial set at instant t0 .
3 Problem Formulation
Let the observer (the 1st player) begin their control actions at the instant t0 > 0, t0 < T , and
u(t) ∈ P ⊂ Rp ,
{ this) and u ≡ 0 on [0, t0 ), the observer
where P is a compact set. If Assumption 1 holds (from now on, we suppose }
can build the compact ellipsoidal informational set X0 = Xt0 (0, y) = x ∈ Rn : |x − x̂(t0 )|2P (t0 ) ≤ 1 − h(t0 ) at
the instant t0 according to equations (6)–(8). The goal of the observer is to minimize the functional
γ(XT (u, y)), (11)
where γ(·) is a non-negative continuous function defined on all compact sets in Rn . The continuity is understood
in the sense of Hausdorff’s convergence.
√ For example, γ(X) = diam X = maxx,y∈X |x−y| or γ(X) = maxx∈X |x|.
In our case, diam XT (u, y) = 2 (1 − h(T ))|P −1 (T )|, where |P | = max
{ √ } |P x|, does not depend on control
|x|≤1
′
u(·). The value maxx∈XT (u,y) |x| = max|l|≤1 l x̂(T ) + 1 − h(T )|l|P −1 (T ) .
It can turn out that h(t0 ) = 1. In this case the resource of disturbances is exhausted and we have v(t) = 0 for
t ≥ t0 . If so, we obtain x(t0 ) = x̂(t0 ) and y(t) ≡ G(t)x̂(t) for t ≥ t0 . Our problem reduces to the ordinary control
one for a linear system (9) with complete information on state vector. If h(t0 ) < 1, we obtain more complicated
situation. It is easily seen from equations (9), (10), that the evolution of informational set Xt (u, y) depends only
on the control u(·) and the innovation function w(t) = y(t) − G(t)x̂(t) that satisfies the LQ-constraints
∫ T
|w(t)|2C(t) ≤ 1 − h(t0 ). (12)
t0
54
We introduce the 2nd player (an opponent) who tries to maximize the functional (11) choosing the function w(t)
under constraints (12). Thus, we have a differential game for equations
˙ 2
x̂(t) = A(t)x̂(t) + B(t)u(t) + C(t)w(t), ν̇(t) = − w(t) C(t) , (13)
(
where ν(t) = 1 − h(t) ≥ 0, C(t) = c(t)C ′ (t) + G(t)P −1 (t))′ C(t). The initial states x̂(t0 ), ν(t0 ) = 1 − h(t0 ),
√
for equations (13) are known from (6)–(8). Note that max diam XT (u, y) = 2 ν(t0 )|P −1 (T )| is obtained under
w(t) = 0 if t ≥ t0 . In the common case, we can consider that
γ(XT (u, y)) = σ(x̂(T ), ν(T )), (14)
where σ(·, ·) is a continuous bounded function. Introduce the vector z(t) = [x̂(t); ν(t)] ∈ Rn+1 , t ∈ [t0 , T ],
ν(t) ∈ [0, 1]. The pair {t, z(t)} is called a position of the game at the instant t. Let us rewrite system (13) once
more:
ż(t) = A(t)z(t) + B(t)u(t) + g(t, w(t)), (15)
[ ]
where A(t) = [A(t), 0; 01×n , 0], B(t) = [B(t); 01×p ], g(t, w) = C(t)w; −|w|2C(t) . Hereafter we use the standard
notation from Matlab, where [A1 , . . . , Ak ] means the row-concatenation of matrices of appropriate dimensions
(sometimes, the comma is replaced by the blank), and [A1 ; . . . ; Ak ] means the column-concatenation.
Let us remark yet that the function σ(z) in terminal functional (14) may be considered Lipschitzean on Rn+1 .
It is fulfilled in many interesting cases.
4 A Solution with HJBI Equation
A strategies of players are defined as arbitrary functions u(t, z) ∈ P and w(t, z, u) with the condition ν ∈ [0, 1].
For any admissible position {t∗ , z∗ }, where ν∗ ∈ (0, 1], the players independently choose the different partitions
∆u and ∆w = {t∗ = t1 < · · · < tk+1 = T }. Let any Borelean control u(·) ∈ P be already fixed. Then the motion
is defined for the interval [ti , ti+1 ) from the partition ∆w as the solution of the stepwise differential equation (see
eq. (15))
ż(t) = A(t)z(t) + B(t)u(t) + g(t, w(ti , z(ti ), u(ti ))).
If for some index i∗ ∈ 1 : k there is an instant t∗ ∈ [ti∗ , ti∗ +1 ) such that ν(t∗ ) = 0, then the resource of 2nd
player is exhausted, and we set w(t) = 0 for t ∈ [t∗ , T ]. On the other hand, if any Borelean function w(·) such
that ν(t) ≥ 0, ∀t ∈ [t∗ , T ], is fixed, then the motion is defined for the interval [ti , ti+1 ) from the partition ∆u as
the solution of the stepwise differential equation
ż(t) = A(t)z(t) + B(t)u(ti , z(ti )) + g(t, w(t)).
For strategies u(·), w(·), and a position {t∗ , z∗ }, introduce the values
c(u(·), t∗ , z∗ ) = lim sup sup σ(z(T )), c(w(·), t∗ , z∗ ) = lim inf inf σ(z(T )), (16)
δ→0 ∆u w[·] δ→0 ∆w u[·]
where δ = |∆u | or δ = |∆w | and |∆| is a diameter of the partition ∆. The supremum in (16) is taken over all
admissible realizations w[·] or u[·]. Strategies u0 (·), w0 (·) are called optimal if
c(u0 (·), t∗ , z∗ ) = inf c(u(·), t∗ , z∗ ), c(w0 (·), t∗ , z∗ ) = sup c(w(·), t∗ , z∗ ),
u(·) w(·)
for any position {t∗ , z∗ }. In work [Lokshin, 1992], it is proved the existence of the optimal strategies u0 (·), w0 (·)
such that
c(u0 (·), t∗ , z∗ ) = c(w0 (·), t∗ , z∗ ) = c0 (t∗ , z∗ )
for any position {t∗ , z∗ }. Here c0 (t∗ , z∗ ) is the value of the game and the pair {u0 (·), w0 (·)} is a saddle point.
In [Lokshin, 1992], a method of computing of the value c0 (t∗ , z∗ ) is developed which based on a recurrent
construction of concave envelopes for functions in auxiliary designs. In our paper, we use the method of HJBI
equation. In addition, note that couterstrategies of 2nd player can be replaced by pure strategies with the same
result.
55
For our problem we need to compute the value of the game c(t0 , z0 ), where from now on, we omit the symbol
0 near c. Fix some N > 0 and temporarily impose an additional constraint on the function w(t): |w(t)| ≤ N .
The corresponding optimal strategies and the value of the game are denoted by u0N (·), wN 0
(·), and cN (t, z),
respectively. For our temporarily case, in [Subbotin, 1999, Theorem 9.1] it was proved that the value of the
game equals cN (t0 , z0 ), where the function cN : [t0 , T ] × Rn × [0, ν0 ] → R satisfies (in corresponding minimax
formalization) the equation
∂cN (t, z)/∂t + HN (t, z, DcN ) = 0 (17)
with boundary condition cN (T, z) = σ(z), ν ∈ [0, ν0 ], as ν(T ) ≤ ν0 . In equation (17) the symbol Dc means the
generalized gradient of function c(t, z) with respect to variables z, and the Hamiltonian HN is defined by the
following way
HN (t, z, l) = min max h(t, z, u, w̃, l), l = [l1 ; l2 ], w̃ = wχ(ν), (18)
u∈P |w̃|≤N
( )
where h(t, z, u, w̃, l) = l1′ (A(t)x̂ + B(t)u) + l1′ C(t)w̃ − l2 |w̃|2C(t) ; the Heaviside function χ(ν) = 1, if ν > 0, and
χ(ν) = 0 otherwise. From now on, the function w̃(t) is called the control with switch. It is easily seen that
lim HN (t, z, l) = H(t, z, l) = l1′ A(t)x̂ + min l1′ B(t)u + |l1 |2C′ (t)C −1 (t)C(t) /(4l2 )
N →∞ u∈P
if ν > 0 and l2 > 0. For equation (17) it is known that its solution in minimax sense coincides with viscosity
solution (see [Subbotin, 1999, Fleming & Soner, 2006]). Note that both the solutions are unique. If the function
cN (t, z) has been built, the optimal strategies of 1-st and 2-nd players are defined as selectors of inclusions
u0N (t, z) ∈ Argmin max h(t, z, u, w̃, DcN (t, z)),
u∈P |w̃|≤N
wN (t, z, u) ∈ Argmax h(t, z, u, w̃, DcN (t, z)).
0
|w̃|≤N
With the help of this strategies we can built a stepwise approximate trajectory until the instant t∗ , where the
resource of the 2nd player is exhausted. After this instant the function w(t) = 0.
Unfortunately, we cannot assert the uniqueness and existence of solution in any sense if the function HN in
(17) is replaced with H. But in [Lokshin, 1992] it is shown that the limit
lim cN (t, z) = c(t, z) (19)
N →∞
exists for any position {t, z} ∈ [t0 , T ] × Rn × [0, 1] uniformly on {t, z}. Thus, equations (17), (18) and relation
(19) gives a method of computing of the value of the game.
4.1 A Numerical Solution
A numerical procedure can be built on the base of [Souganidis, 1985, Souganidis, 1999, Fleming & Soner, 2006].
For technical reason the all the matrices in (1), (2) will be considered Lipschitzean in t. Then the left-hand side
of equation (15) F (t, z, u, w) = A(t)z + B(t)u + g(t, w), where u ∈ P, |w| ≤ N , satisfies the uniform Lipschitz
condition |F (t1 , z1 , u, w) − F (t2 , z2 , u, w)| ≤ C1 (|z1 − z2 | + |t1 − t2 |). In addition, the function σ(z) from (14) is
Lipshitzean, i.e. |σ(z1 ) − σ(z2 )| ≤ C2 |z1 − z2 |.
For approximation of cN (t, z), we consider the partition ∆ = {t0 < t1 < · · · < tK(∆)+1 = T } of the
interval [t0 , T ]. The diameter maxi |ti+1 − ti | of the partition is denoted by |∆| as before. Define the function
c∆ : [0, T ] × Rn × [0, ν0 ] → R as
{ }
c∆ (T, z) = σ(z) on Rn × [0, ν0 ], c∆ (t, z) = min max c∆ (ti+1 , z + (ti+1 − t)F (t, z, u, w̃) ,
u∈P |w̃|≤N
(20)
if t ∈ [ti , ti+1 ), and i ∈ 0 : K(∆).
We do the computation backward from right to the left. For i = K(∆) we have
{ }
c∆ (t, z) = min max σ(z + (T − t)F (t, z, u, w̃) , t ∈ [tK(∆) , T ).
u∈P |w̃|≤N
Using [Souganidis, 1999, Theorem 4.4], we obtain
56
Theorem 1 Under |∆| → 0, the function (20) converges to cN (t, z) locally uniformly on [t0 , T ] × Rn × [0, ν0 ].
The function cN (t, z) is the unique solution of equation (17) in minimax or viscosity sense. Besides, there exists
a constant L(C2 , ||σ||, ||σz ||), such that |c∆ (t, z) − cN (t, z)| ≤ L|∆|1/2 for all (t, z) ∈ [t0 , T ] × Rn × [0, ν0 ]. Here
|| · || is the sup-norm of corresponding function.
Let us introduce the attainability set W(t) of system (13) with support function
∫ t √ ∫
t
ρ(l|W(t)) = max l′ x = ρ(l|X(t, t0 )X0 ) + ρ(l|B(s)P)ds + ν0 |C′ (s)l|2 ds,
x∈W(t) t0 t0
and the union W = ∪t∈[t0 ,T ] W(t). The set W is compact. We can suggest the following
Numerical Algorithm
1. Choose a finite set (a grid) N = {zk } ⊂ W × [0, ν0 ], k ∈ 1 : K1 .
2. Select a partition ∆ = {t0 < t1 < · · · < tK+1 = T } of [0, T ].
( )
3. Form and remember the function σK (z) = min max σ z + (tK+1 − tK )F (tK+1 , z, u, w̃) and corresponding
u∈P |w̃|≤N
optimal controls u∗K and w̃K
∗
(u), where z ∈ N .
( )
4. On subsequent steps the grid function is formed: σi (z) = min max σi+1 z + (ti+1 − ti )F (ti+1 , z, u, w̃) and
u∈P |w̃|≤N
corresponding optimal controls u∗i and w̃i∗ (u). If the value z + (ti+1 − ti )F (ti+1 , z, u, w̃) does not lie in the
grid, then this value is changed for the nearest element from N .
5. The value σ0 (z) gives an approximate value of the game.
Before the calculation of σ0 (z) we need to choose the sufficiently large number N .
5 An Example
Let us consider the rectilinear movement of an airplane in the vertical plane at h height: x1(nom) = L + V t,
x2(nom) = h. The real initial state [x̃0 ; ỹ0 ] may be differ from [L; h] and unknown. The deviation from the basic
movement is described by the system
ẋ1 = x3 , ẋ3 = u1 , ẋ2 = x4 , ẋ4 = u2 , t ∈ [0, T ],
where x1 = x̃1 − x1(nom) , x2 = x̃2 − x2(nom) , x3 = x̃3 − V , x4 = x̃4 . The control accelerations are limited by the
√
constraint u21 + u22 ≤ 10. The model of measurements is of the form ỹ = (L − x̃1 )2 + x̃22 . Linearizing this with
respect to basic movement, we obtain
√ √
y(t) = g1 (t)x1 + g2 (t)x2 + v(t), g1 (t) = V t/ V 2 t2 + h2 , g2 (t) = h/ V 2 t2 + h2 ,
∫T
where v(t) is a disturbance with the constraint 0 v 2 (t)dt ≤ 1. It is possible to check that the system is
completely observable. Therefore, for any 0 < t0 < T the initial informational set X0 is bounded. Given
numerical data: L = 2500 m, h = 10000 m, V = 1200√m/s, T = 30 s, we calculate the set X0 for t0 = 10 s,
y(t) = g1 (t)(−500 + 200t) + g2 (t)(−1000 + 10t) + sin t/ 20.
Here A = [0, 0, 1, 0; 0, 0, 0, 1; 0, 0, 0, 0; 0, 0, 0, 0], B = [0, 0; 0, 0; 1, 0; 0, 1], C(t) = 04×1 , G(t) = [g1 (t), g2 (t), 0, 0],
c(t) = 1. Let us write system (6)–(8):
Ṗ (t) = −P (t)A − A′ P (t) + G′ (t)G(t), P (0) = 0;
˙ = −A′ d(t) + G′ (t)y(t) + P (t)Bu(t), d(0) = 0;
d(t)
q̇(t) = y 2 (t) + 2d′ (t)Bu(t), q(0) = 0.
For t0 = 10 we get x̂0 ≈ [2331; −1898; 200; −89], h(t0 ) = 0.3574. The diameter of initial set equals 29.0228. To
the end of time interval it increases to 143.2507 and does not
{ depend√of the control of 1st} player. Of course, the
control of 2nd player must be zero. Let σ(z) = max|l|≤1 l′ x̂(T ) + 1 − h(T )|l|P −1 (T ) , i.e. we minimize the
maximal deviation. We have P −1 (T ) = [1003, −3594, 1, −120; −3594, 12902, −2, 431; 1, −2, 0, 0; −120, 431, 0, 14].
Using Numerical Algorithm and setting ν0 = 0, we obtain the value of the game as c(t0 , z0 ) = 71.3471. More
detail consideration of this example will be done in subsequent works.
57
6 Conclusion
A control problem for non-stationary linear systems is considered. These systems as a rule describe the deviation
from a rated trajectory. According to the output a controller builds the informational set containing the real
state vector and tries to minimize the functional of final informational set. A 2nd player using disturbances tries
to disturb the 1st player and maximizes the functional. The situation is reduced to a differential game with
complete information for linear differential equation describing the evolution of center of informational ellipsoid.
To solve this problem we use the theory of differential games. An approximation of the value of the game may
be found by integration of corresponding HJBI equation, the solution of which is understood in a generalized
sense. The optimal strategies are also defined due to this solution. The numerical approximation is specified and
the estimation of the rate of convergence is given for the approximating scheme. An example with a rectilinear
movement of an airplane is examined.
Acknowledgements
This work was supported by Russian Science Foundation (RSF) under Project No. 16-11-10146.
References
[Kurzhanski & Varaiya, 2014] Kurzhanski, A., & Varaiya, P. (2014). Dynamics and Control of Trajectory Tubes:
Theory and Computation. SCFA. Birkhäuser, Boston.
[Krasovskii, 1985] Krasovskii, N.N. (1985). Control of Dynamic Systems. Nauka, Moscow. (in Russian)
[Lokshin, 1992] Lokshin, M.D. (1992). Differential games with integral constraints on the control actions. Dif-
ferential Equations, 28(11), 1602-1610. (translation from Russian)
[Souganidis, 1985] Souganidis, P.E. (1985). Approximation Schemes for Viscosity Solutions of Hamilton-Jacobi
Equations. J. Differen. Equat., vol. 59, 1-43.
[Souganidis, 1999] Souganidis, P.E. (1999). Two-Player, Zero-Sum Differential Games and Viscosity Solutions.
In Annals of the Intern. Society of Dynamic Games (ed. Tamer Basar), vol. 4. Birkhüser, Boston, pp.
69–104.
[Subbotin, 1999] Subbotin, A.I. (1999). Constructive Theory of Positional Differential Games and Generalized
Solutions to Hamilton-Jacobi Equations. In Annals of the Intern. Society of Dynamic Games (ed. Tamer
Basar), vol. 4. Birkhüser, Boston, pp. 3–67.
[Ananyev, 2017] Ananyev, B.I. (2017). Parameter Optimization for Estimation of Linear Non-Stationary Sys-
tems. In 8th Intern. Scientific Conf. on Physics and Control. Physcon’2017. Florence, Italy. July 17-19,
2017. Internet publ. (to be published)
[Ananiev, 2011] Ananiev, B.I. (2011). Some Problems of Observations’ Control. American Institute of Physics.
Conference Proceedings. Vol. 1404, No. 1, pp. 235-241. Melville, NY 11747-4502, USA.
[Ananyev, 2012] Ananyev, B.I. (2012). American Institute of Physics. Conference Proceedings. Vol. 1487, No.
1, pp. 143-150.
[Fleming & Soner, 2006] Fleming, W.H. & Soner, H.M. (2006). Controlled Markov Processes and Viscosity So-
lutions. Springer-Verlag, New York.
58