Abstract
In differential games the a posteriori analysis of motions, namely, trajectories of the dynamics and the analysis of the players’ controls generating these trajectories are very important. This paper is devoted to solving problems of reconstruction of trajectories and controls in differential games using known history of inaccurate measurements of a realized trajectory. A new method for solving reconstruction problems is suggested and justified for a class of differential games with dynamics, linear in controls and non-linear in state coordinates. This method relies on necessary optimality conditions in auxiliary variational problems. An illustrating example is exposed.
Access provided by CONRICYT-eBooks. Download chapter PDF
Similar content being viewed by others
7.1 Introduction
This paper is devoted to solving inverse problems of reconstruction of players’ trajectories and controls in differential games, using known inaccurate measurements of the realized trajectories. The a posteriori analysis is an important part of the decision making in the future. Inverse problems may occur in many areas such as economics, engineering, medicine and many others that involve the task of reconstruction of the players’ controls by known inaccurate trajectory measurements.
The inverse problems have been studied by many authors. The approach suggested by Osipov and Kryazhimskii [6, 7] is one of the closest to the material of this paper. The method suggested by them reconstructs the controls by using a regularized (a variation of Tikhonov regularization [12]) procedure of control with a guide. This procedure allows to reconstruct the controls on-line. It is originated from the works of Krasovskii’s school on the theory of optimal feedback [3, 4].
Another method for solving dynamic reconstruction problems by known history of inaccurate measurements has been suggested by Subbotina et al. [10]. It is based on a method, which use necessary optimality conditions for auxiliary optimal control problems [9]. This method has been also developed in [5, 8, 10, 11]. A modification of this approach is presented in this paper. It relies on necessary optimality conditions in an auxiliary variational problem on extremum for an integral functional. The functional is a variation of a Tikhonov regularizator.
In this paper the suggested method is justified for a special class of differential games with dynamics linear in controls and non-linear in state coordinates. Results of simulation are exposed.
7.2 Dynamics
We consider a differential game with dynamics of the form
Here G(x, t) is an n × n matrix with elements g ij(x, t) : R n × [0, T] → R, i = 1, ..., n, j = 1, ..., n that have continuous derivatives
In (7.1) x i(t) is the state of the ith player, while u i(t) is the control of the ith player, restricted by constraints
We consider piecewise continuous controls with finite number of points of discontinuity.
7.3 Input Data
It is supposed that some base trajectory x ∗(⋅) : [0, T] → R n of system (7.1) has been realized on the interval t ∈ [0, T]. Let u ∗(⋅) : [0, T] → R n be the piecewise continuous control satisfying constrains (7.2) that generated this trajectory.
We assume that measurements y δ(⋅, δ) = y δ(⋅) : [0, T] → R n of the base trajectory x ∗(t) are known and they are twice continuously differentiable functions that determine x ∗(t) with the known accuracy δ > 0, i.e.
7.4 Hypotheses
We introduce two hypotheses on the input data.
Hypothesis 7.1
There exist such compact set Ψ ⊂ R n , such constant r > 0 and such constants \( \underline \omega >0,\ \overline \omega >0,\ \omega '>0\) that
Let’s introduce the following constants
which will be used in Hypothesis 7.2 and Theorem 7.1.
Hypothesis 7.2
There exist such constants \(\delta _0\in (0,\min \{0.5 r,\frac {1}{R_w}\}]\) and \(\overline Y>0\) that for any δ ∈ (0, δ 0]
and for any δ ∈ (0, δ 0] exists such compact Ω δ ⊂ [0, T] with measure \(\mu \varOmega ^\delta = \beta ^\delta \stackrel {\delta \to 0}{\longrightarrow }\) that
Remark 7.1
Conditions (7.6) reflect the fact that the right hand sides of Eq. (7.1) are restricted.
Remark 7.2
In Hypothesis 7.2 the constant \(\overline Y\) is unified for all inequalities to simplify the further calculations and explanations.
Remark 7.3
Hypothesis 7.2 allows the functions \(\dot y^\delta (\cdot )\) to be able to approximate piecewise continuous functions \(\dot x^*(\cdot ) = g(x^*(\cdot ),\cdot )u^*(\cdot )\).
7.5 Problem Statement
Let’s consider the following reconstruction problem: for a given δ ∈ (0, δ 0] and a given measurement function y δ(⋅) fulfilling estimates (7.3) and Hypothesis 7.2 to find a function u(⋅, δ) = u δ(⋅) : [0, T] → R n that satisfies the following conditions:
-
1.
The function u δ(⋅) belongs to the set of admissible controls, i.e. the set of piecewise continuous functions with finite number of points of discontinuity satisfying constraints (7.2);
-
2.
The control u δ(⋅) generates trajectory x(⋅, δ) = x δ(⋅) : [0, T] → R n of system (7.1) with boundary condition x δ(T) = y δ(T). In other words, there exists a unique solution x δ(⋅) : [0, T] → R n of the system
$$\displaystyle \begin{aligned} \dot x^\delta(t) = G(x^\delta(t),t)u^\delta(t),\quad t\in [0,T] \end{aligned}$$that satisfy the boundary condition x δ(T) = y δ(T).
-
3.
Functions x δ(⋅) and u δ(⋅) satisfy conditions
$$\displaystyle \begin{aligned} \lim\limits_{\delta\rightarrow 0}\|x^\delta_i(\cdot) - x^*_i(\cdot)\|{}_{C_{[0,T]}} = 0,\quad \lim\limits_{\delta\rightarrow 0}\|u^\delta_i(\cdot) - u^*_i(\cdot)\|{}_{L_{2,[0,T]}} = 0,\quad i = 1,...,n. \end{aligned} $$(7.8)
Hereinafter
is the norm in the space of continuous functions C and
is the norm in space L 2.
7.6 A Solution of the Inverse Problem
7.6.1 Auxiliary Problem
To solve the inverse problem in Sect. 7.5, we introduce an auxiliary variational problem (AVP) for fixed parameters δ ∈ (0, δ 0], α > 0 and a given measurement function y δ(⋅) satisfying estimates (7.3) and Hypothesis 7.2.
We consider the set of pairs of continuously differentiable functions F xu = {{x(⋅), u(⋅)} : x(⋅) : [0, T] → R n, u(⋅) : [0, T] → R n} that satisfy differential equations (7.1) and the following boundary conditions
Hereinafter G −1 is the inverse matrix for non degenerate matrix G. Let us remark that due to Hypothesis 7.1, the inverse matrix G −1(y δ(T), T) exists.
AVP is to find a pair of functions x(⋅, δ, α) = x δ, α(⋅) : [0, T] → R n and u(⋅, δ, α) = u δ, α(⋅) : [0, T] → R n such that {x δ, α(⋅), u δ, α(⋅)}∈ F xu and such that they provide an extremum for the integral functional
Here α is a small regularising parameter [12] and \(\|f\| = \sqrt {\sum \limits _{i=1}^n f_i^2},\ f\in R^n\) is Euclidean norm in R n.
7.6.2 Necessary Optimality Conditions in the AVP
We can write the necessary optimality conditions for the AVP (7.1), (7.10), (7.9) in Lagrange form [14]. Lagrangian for the AVP has the form
where λ(t) : [0, T] → R n is the Lagrange multipliers vector.
The 2n corresponding Euler equations are
The first n equations in (7.11) can be rewritten in vector form:
Hereinafter ‹a, b› means the scalar product of vectors a ∈ R n, b ∈ R n and \(\displaystyle \frac {\partial G}{\partial x_i}(x(t),t)\) is a matrix with elements \(\displaystyle \frac {\partial g_{jk}}{\partial x_i}(x(t),t)\), j = 1, ..., n, k = 1, ..., n.
The last n equations in (7.11) define the relations between the controls u i(t) and the Lagrange multipliers λ i(t), i = 1, …, n:
Hereinafter G T means transpose of a matrix G.
We can substitute equations (7.13) into (7.12) and (7.1) to rewrite them in the form of Hamiltonian equations, where the vector s(t) = −λ(t) plays the role of the adjoint variables vector:
By substituting (7.13) into (7.9), one can obtain boundary conditions, written for system (7.14):
Thus, we have got the necessary optimality conditions for the AVP (7.1), (7.10), (7.9) in Hamiltonian form (7.14), (7.15).
7.6.3 A Solution of the Reconstruction Problem
Let’s introduce the function
where x δ, α(⋅), s δ, α(⋅) are the solutions of system (7.14) with boundary conditions (7.15).
We now introduce the cut-off functions
We consider the functions \(\hat u_i^\delta (\cdot )\) as the solutions of the inverse problem described in Sect. 7.5. We choose α = α(δ) in a such way that \(\alpha (\delta )\stackrel {\delta \to 0}{\longrightarrow } 0\).
7.6.4 Convergence of the Solution
In this paper a justification for the suggested method is presented for one sub-class of considered differential games (7.1), (7.2). Namely, we consider from now dynamics of form (7.1), where matrixes G(x, t) are diagonal with non-zero elements on the diagonals. The dynamics in such case have the form
where the functions g i(x, t) = g ii(x, t), i = 1, …, n are the elements on the diagonal of the matrix G(x, t).
Condition \( \underline \omega ^2\leq | det G(x,t)|\leq \overline \omega ^2\) in Hypothesis 7.1 in such case is replaced by equal condition
Necessary optimality conditions (7.14) has now the form
with boundary conditions
The following lemma is true.
Lemma 7.1
For δ ∈ (0, δ 0] twice continuously differentiable measurement functions \(y_i^\delta (\cdot ),\ i=1,\ldots ,n\) satisfying estimates (7.3) and Hypothesis 7.2 fulfill the following relations
Proof
The first relation in (7.21) is true due to (7.3). Let’s prove the second one.
Relying upon Luzin’s theorem [2] one can find for the piecewise continuous function u ∗(⋅) such constant \(\overline Y^u\) that for any δ ∈ (0, δ 0] and all i = 1, …, n there exist such twice continuously differentiable functions \(\overline u^\delta _i(\cdot ):[0,T]\to R\) and such set \(\varOmega ^\delta _u\subset R\) with measure \(\mu \varOmega ^\delta _u = \beta ^\delta _u\) that
Let’s estimate the following expression first (hereinafter in the proof i = 1, …, n).
The integral in (7.23) can be calculated by parts.
To estimate the whole expression (7.24) we first estimate the difference \(\mathbf {V} = y^\delta _i(t) - x_i^*(0) - \int \limits _0^t \overline u^\delta _i(\tau )g_i(y^\delta (\tau ),\tau )d\tau \). In order to do this, we estimate integral
where set \(\varOmega ^t_{\geq \delta }=\{\tau \in [0,t]:|\overline u^\delta _i(\tau ) - u_i^*(\tau )|\geq \delta \}\) and set \(\varOmega ^t_{<\delta }=\{\tau \in [0,t]:|\overline u^\delta _i(\tau ) - u_i^*(\tau )|<\delta \}\).
The first term in (7.25)
Remark 7.4
Let’s remember that hereinafter when the first argument of functions g i(x, t), i = 1, …, n belongs to compact Ψ from Hypothesis 7.1, relations (7.18) are true.
Using (7.22), the second term in (7.25)
From (7.25), (7.26) and (7.27) follows that
We can now estimate function V in (7.24):
Thus, the term \(\mathbf {UV}|{ }_0^T\) in sum (7.24) can be estimated as
Using (7.7), (7.22) and (7.29), the term \(\int \limits _0^T \mathbf {VdU}dt\) in (7.24) can be estimated in the following way
Combining estimates (7.30) and (7.31), we can now estimate expression (7.24).
Finally, we can use the first mean value theorem for definite integrals and estimate (7.32) to get
It follows from (7.33) that
Remember that we consider such function \(\overline u^\delta _i (\cdot )\) that \( \lim \limits _{\delta \rightarrow 0}\left \|\overline u^\delta _i(\cdot ) - u^*_i(\cdot )\right \|{ }_{L_{2,[0,T]}} = 0. \) So, from the triangle inequality \(\|f_1(\cdot ) + f_2(\cdot )\|{ }_{L_{2,[0,T]}}\leq \|f_1(\cdot )\|{ }_{L_{2,[0,T]}} + \|f_2(\cdot )\|{ }_{L_{2,[0,T]}}\) follows that
which was to be proved. □
Theorem 7.1
For any fixed δ ∈ (0, δ 0] there exists such parameter \(\alpha ^\delta _0 = \alpha ^\delta _0(\delta )\) that the solution \(x^{\delta ,\alpha ^\delta _0}(\cdot )\) , \(s^{\delta ,\alpha ^\delta _0}(\cdot )\) of system (7.19) with boundary conditions (7.20) is extendable and unique on t ∈ [0, T].
Moreover, \(\lim \limits _{\delta \to 0}\alpha ^\delta _0(\delta ) = 0\) and
where
Proof
Let’s introduce new variables:
Their derivatives are
System (7.19) can be rewritten in this variables as
where
Boundary conditions (7.20) in new variables take the form
As it follows from Hypothesis 7.1, the right hand side of system (7.38) is locally Lipschitz on Ψ × [0, T]—so, by Cauchy theorem there exists such interval [T 0, T] ⊂ [0, T] that solutions z δ, α(⋅) : [T 0, T] → R n, w δ, α(⋅) : [T 0, T] → R n of system (7.38) with boundary conditions (7.40) exist and are unique on t ∈ [T 0, T]. Moreover, due to continuity of the solutions and zero boundary conditions (7.40), there exists such interval [t 1, T] ⊂ [T 0, T] that
where the constant R w is defined in (7.5).
Let’s now extend the solution further in reverse time (to the left from t 1 on time axis). As the solution is continuous, we can always extend it up to such moment t 0 that either \(z^{\delta ,\alpha }_i(t_0)=2\alpha \delta R_w,\) i ∈{1, …, n} or \({w^{\delta ,\alpha }_i(t_0)=2\alpha ^2\delta R_w,\ i\in \{1,\ldots ,n\}}\) or extend it up to t = 0. If we are able to extend it up to t = 0 without reaching values 2αδR w, 2α 2 δR w (the second case), then
In the first case there exists such moment t 0 ∈ [0, T] that
Let’s consider this case closer.
We introduce a new system of ODEs for functions \(\overline z_i(\cdot )\), \(\overline w_i(\cdot ),\ i = 1,\ldots ,n\)
with boundary conditions
where \(z^{\delta ,\alpha }_i(t)\), \(w^{\delta ,\alpha }_i(t)\) are solutions of system (7.38) with boundary conditions (7.40), constrained by (7.41).
System (7.42) is a heterogeneous linear system of ODEs with time-dependent coefficients, continuous on t ∈ [t 0, T]. So, the solution of (7.42), (7.43) exists and is unique on t ∈ [t 0, T].
Let’s now prove that the solutions of (7.42), (7.43) coincide with the solutions of (7.38), (7.40). To do this, we introduce residuals
Subtracting Eq. (7.42) from (7.38) (with substituted solutions z δ, α(t), w δ, α(t)), we get
with boundary conditions
As a homogenous system of linear ODEs with continuous time-dependent coefficients, system (7.44) with zero boundary conditions has the only trivial solution
That means that \(z^{\delta ,\alpha }(t) = \overline z(t)\), \(w^{\delta ,\alpha }(t) = \overline w(t),\ t\in [t_0,T]\).
Now let’s study the properties of the solutions \(\overline z(t)\), \(\overline w(t)\) of system (7.42) with boundary conditions (7.43). System (7.42) can be rewritten in vector form
where
and the 2n × 2n matrix A(t) can be written in the block form \(A(t) = \left ( \begin {array}{cc} O & G_A(x,t) \\ I_n & O \\ \end {array} \right )\), where I n is an identity matrix, O is an n × n zero matrix,
Solutions of system (7.42) can be written in the following form with the help of Cauchy formula for solutions of a heterogenous system of linear ODEs with time-dependent coefficients. One can easily check that for boundary conditions, given at the point t = T (instead of t = 0), it has the form
were Φ(⋅) is an n × n fundamental matrix of solutions for the homogenous part of system (7.42). This matrix can be chosen as
One can check that after expanding the kth powers in the sum in the latter formula and folding the sum again, using the Taylor series for sin and cos functions, we can get that \(\varPhi (t) = \left ( \begin {array}{cc} \varPhi _1(t) & \varPhi _2(t) \\ \varPhi _3(t) & \varPhi _1(t) \\ \end {array} \right )\), where Φ 1(t), Φ 2(t), Φ 3(t) are diagonal matrixes with ith elements on diagonals
where continuous function
Using (7.18), one can obtain that
Due to simple structure of matrix Φ(t), one can check that inverse matrix \(\varPhi ^{ - 1}(t) = \left ( \begin {array}{cc} \varPhi _1(t) & - \varPhi _2(t) \\ - \varPhi _3(t) & \varPhi _1(t) \\ \end {array} \right )\).
Let’s return to (7.49). Here \(Z(T) = \overrightarrow {0}\), so vector \(Z(t) = -\varPhi (t) \int \limits _t^T\varPhi ^{-1}(\tau ) F(\tau )d\tau \) has the following coordinates
To estimate these expressions, we consider the following expression
where function \(f^\delta _i(\cdot )=f^\delta _i(\cdot ,\delta ):[0,T]\to R\) depends on δ and is continuous in the first argument for any δ ∈ (0, δ 0].
Let’s introduce functions \(\varphi _i (\tau ) = \left (\sqrt {(T - \tau )\int \limits _\tau ^T g_i^2(x^{\delta ,\alpha }(\theta ),\theta )d\theta }\right ),\ i=1,\ldots ,n\), which are continuously differentiable in τ.
Note that all following calculations in the proof are true for all i ∈{1, …, n}.
Using Hypothesis 7.1, we can estimate the derivative
So, φ i(τ) is a decreasing function with restricted derivative and φ i(T) = 0. This means that we can construct a finite increasing sequence \(\{\tau _1<\tau _2<\ldots <\tau _{n_{\varphi _i}},\ n_{\varphi _i} \in N\}\) that has the following properties:
as the derivative \(\dot \varphi (t)\) is restricted by (7.56).
Let’s add to this sequence elements τ 0 = t 0 and \(\tau _{(n_{\varphi _i}+1)} = T\).
Integral (7.55) can be rewritten as
Because \(\cos \left (\frac {\varphi _i (\tau )}{\alpha }\right )\) is sign-definite on \(\tau \in [\tau _j,\tau _{j + 1}],\ j = 0,\ldots ,n_{\varphi _i}\) and \(f^\delta _i(\tau )\) is continuous, it follows from the first mean value theorem for definite integrals that for each \(j = 0,\ldots ,n_{\varphi _i}\) there exists such point \(\tilde \tau _j\in [\tau _j,\tau _{j + 1}]\) that \(\int \limits _{\tau _j}^{\tau _{j + 1}} \cos \left (\frac {\varphi _i (\tau )}{\alpha }\right )f^\delta _i(\tau )d\tau = f^\delta _i(\tilde \tau _j)\int \limits _{\tau _j}^{\tau _{j + 1}} \cos \left (\frac {\varphi _i (\tau )}{\alpha }\right )d\tau \). Combining the terms of sum (7.58) by pairs [τ j, τ j+1], [τ j+1, τ j+2], we get
To estimate expression (7.59), we first make the following estimates:
as (T − τ) ≠ 0 for \(j<n_{\varphi _i}\). We can integrate (7.60) by parts.
Here
One can check that the derivative
So, the term \(\mathbf {UV}\big |{ }_{\tau _j}^{\tau _{j+1}}\) in (7.61) can be estimated by using (7.57) and (7.62) as
where the constants R 1, R 2, R 3 are defined in (7.5). Let’s emphasize that these constants don’t depend on δ and α.
Now let’s estimate the term \(\int _{\tau _j}^{\tau _{j+1}} \mathbf {VdU}\) in (7.61).
Applying estimates (7.63) and (7.64) to (7.60)–(7.61), we get
Now let’s return to expression (7.59). By splitting the last integral term in (7.59) as \(\int \limits _{\tau _{j + 1}}^{\tau _{j + 2}} = \int \limits _{\tau _j}^{\tau _{j + 2}} - \int \limits _{\tau _j}^{\tau _{j + 1}}\), we get
By Heine–Cantor theorem, every continuous function defined on a closed interval is uniformly continuous. So, continuous \(f^\delta _i(\tau )\) is uniformly continuous on [t 0, T]. In other words,
Remark 7.5
As \(f^\delta _i(\tau )\) is uniformly continuous on [t 0, T], we are able to choose the same \(\alpha _1^\delta = \alpha _1^\delta (\delta )\) in (7.67) for each \(j=0,\ldots ,(n_{\varphi _i}+1)\) as [τ j, τ j+2] ⊂ [t 0, T].
Combining (7.57), (7.65), (7.66), (7.67), we get
To be specific, let’s assume that the number \(n_{\varphi _i}\) is odd. Then
Using (7.68), let’s first estimate the sum
where \({n_{\varphi _i} \leq \frac {T\overline \omega }{\alpha \underline \omega ^2}}\) and \({\overline f^\delta _i = \max \limits _{\tau \in [t_0,T]}f^\delta _i(\tau )}\).
The following sum can be estimated by substituting the denominator in the fraction with it’s minimal possible value (7.57) and reversing the order of terms in the sum.
The partial sum \(\sum \limits _{j = 1}^{0.5 (n_{\varphi _i}-1)-1}\frac {1}{j}\) of a harmonic series can be estimated by Euler–Mascheroni formula \(\sum \limits _{n=1}^{k}{\frac {1}{n}} \leq (\ln k)+1\). Thus, continuing estimates (7.70) we get
We have estimated the second term of sum in the right hand side of (7.69). Using (7.57), one can get the following relations for the first, third and forth terms in (7.69).
Remark 7.6
We assumed that the number \(n_{\varphi _i}\) is odd. In the case of even \(n_{\varphi _i}\) the calculations are similar, because the only difference is in formula (7.69), where the lower limit of the integral \(\int \limits _{\tau _{n_{\varphi _i}-2}}^{\tau _{n_{\varphi _i}}} \cos \left (\frac {\varphi _i (\tau )}{\alpha }\right )f^\delta _i(\tau )d\tau \) is exchanged for \(\tau _{n_{\varphi _i}-1}\).
Finally, applying (7.71) and (7.72) to (7.69), we get
For any given δ ∈ (0, δ 0] there exists a constant \({\overline f^\delta _i = \overline f^\delta _i(\delta )}\) (as \(f^\delta _i(\tau )\) is continuous for δ ∈ (0, δ 0]). We can always find such parameter \(\alpha _2^\delta = \alpha _2^\delta (\delta )\) that
This is possible because \({\lim \limits _{\alpha \to 0}\alpha |\ln \alpha | = 0}\). Thus, for any
we have
where the constants R 3, R 4 are defined in (7.5).
We can apply this result to expressions (7.54). First, let’s estimate expression
for which \(f^\delta _i(\tau ) = F_i(z^{\delta ,\alpha }(\tau ),w^{\delta ,\alpha }(\tau ),\tau )/\alpha ^2 \stackrel {not}{=} f^\delta _{i,1}(\tau ) \) in the sense of (7.55). It follows from (7.39), (7.41) and Hypotheses 7.2, 7.1 that
For \(\alpha \leq \alpha ^1_0\), where \(\alpha _0^1\) is defined in the same way as α 0 in (7.67), (7.73), (7.74), but assuming \(f^\delta _i(\tau ) = f^\delta _{i,1}(\tau ) \) and \(\overline f^\delta _i(\tau ) = \overline f^\delta _{i,1}(\tau ) \), estimates (7.75) and (7.53) give us
Let’s introduce \(\alpha _0^2\) that is defined in the same way as α 0 in (7.67), (7.73), (7.74), but assuming
One can use the scheme of proof (7.55)–(7.78) and (7.51)–(7.53) to obtain that for \(\alpha \leq \min \{\alpha ^1_0,\alpha ^2_0\}\) the following estimates are true as well
Remark 7.7
Estimates (7.78)–(7.81) are true under combined condition
Combining (7.54) and (7.78)–(7.81), we get
For \(\delta :0<\delta \leq \delta _0\leq \frac {1}{R_w}\), \(\alpha \in (0,\alpha _0^\delta ]\), as far as \(z^{\delta ,\alpha }(t) = \overline z(t)\), \(w^{\delta ,\alpha }(t) = \overline w(t),\ t\in [t_0,T]\),
Remark 7.8
Estimates (7.83) are true for t 0 ∈ [0, T) as long as solutions z δ, α(⋅), w δ, α(⋅) of system (7.38) with boundary conditions (7.40) exist and are unique on t ∈ [t 0, T] and (7.41) is true.
But (7.83) means that for δ ∈ (0, δ 0], \(\alpha \in (0,\alpha _0^\delta ]\) at t = t 0 (in particular)
which is contrary to the assumption that either \({z^{\delta ,\alpha }_i(t_0)=2\alpha \delta R_w,\ i\in \{1,\ldots ,n\}}\) or \({w^{\delta ,\alpha }_i(t_0)=2\alpha ^2\delta R_w,\ i\in \{1,\ldots ,n\}}\). That means that such moment t 0 does not exist.
In other words, we have proved that we can extend the solutions z δ, α(⋅), w δ, α(⋅) up to t = 0 and
for δ ∈ (0, δ 0] and \(\alpha \in (0,\alpha _0^\delta ]\).
As far as we can extend solutions z δ, α(⋅), w δ, α(⋅) on t ∈ [0, T], we can return to variables (7.36)
Applying the result (7.84) (see Remark 7.8), we get that
for δ ≤ (0, δ 0] and \(\alpha \in (0,\alpha _0^\delta ]\).
It follow from (7.85) and Hypothesis 7.2 that
which means that
Let’s now make the following calculations.
for δ ≤ (0, δ 0] and \(\alpha \in (0,\alpha _0^\delta ]\).
It follows from (7.87) and (7.85) that
Relation (7.88) and Lemma 1 imply that
which was to be proved. □
Let’s now consider for a fixed δ ∈ (0, δ 0] cut-off functions
where the functions \(u_i^{\delta ,\alpha ^\delta _0}(\cdot ),\ i=1,\ldots ,n\) are defined in (7.35) and \(\alpha ^\delta _0\) is introduced in Theorem 7.1 in (7.82).
It follows from Theorem 7.1 that
Combining (7.89) and constraints (7.2), we get
Since all terms in the last expression in (7.90) are non-negative, we obtain
Now let’s prove the following lemma
Lemma 7.2
The system of differential equations
where \(\hat u_i^\delta (\cdot )\) is defined in (7.89) for a fixed δ ≤ (0, δ 0], have a unique solution \(x(\cdot ) \stackrel {not}{=} \hat x^\delta (\cdot ):[0,T]\to R^n\) . Moreover,
Proof
Let’s introduce new variables
where \(x^{\delta ,\alpha ^\delta _0}(t)\) is the solution of system (7.19) with boundary conditions (7.20).
System (7.93) in this variables has the form
The right-hand sides of this equations
Since estimates (7.95) are true and the function \(|\hat u^\delta _i(\cdot ) - u^{\delta ,\alpha ^\delta _0}_i(\cdot )|\) is continuous, the solution of system (7.94) is unique and can be extended on [0, T] [13]. Thus, the solutions \(\hat x^\delta _i(t) = \triangle x_i(t) - x^{\delta ,\alpha ^\delta _0}_i(t),\ i=1,\ldots ,n\) of system (7.93) can be extended on t ∈ [0, T] as well.
From (7.95) it follows that
Hence,
Applying the Grönwall–Bellman inequality, we get
Here \(\|\triangle x(T)\| \leq \sqrt {n}\delta \stackrel {\delta \to 0}{\longrightarrow } 0\). Since (7.91), \(\int \limits _t^T |\hat u^\delta _i(\tau ) - u^{\delta ,\alpha ^\delta _0}_i(\tau )|d\tau \stackrel {\delta \to 0}{\longrightarrow } 0,\ t\in [0,T]\). So, \(\|\triangle x(t)\|\stackrel {\delta \to 0}{\longrightarrow } 0,\ t\in [0,T]\). In other words,
Combining this result with result of Theorem 7.1 (7.34), we get
which was to be proved.
Lemma 7.2, definition (7.89) and formula (7.92) mean that functions (7.89) can be considered as solution of the inverse problem described in Sect. 7.5.
7.7 Remarks on the Suggested Method
Note that Hypotheses 7.2 and 7.1, Theorem 7.1 and Lemmas 7.1 and 7.2 provide that in case of diagonal matrix G(x, t) the solution for the inverse problem described in Sect. 7.5 can be found as
The case of non diagonal non degenerate matrix G(x, t) is more interesting. In this case the solution can still be found by inversing the matrix G(y δ(t), t)
but it involves finding the inverse matrix G −1(y δ(t), t) for each t ∈ [0, T].
One can modify the algorithm suggested in Sect. 7.6 to solve the inverse problem for the case of non-diagonal matrix G(y δ(t), t) as well. The justification uses the same scheme of proof, but is more complex due to more complicated form of system (7.19). It will be published in later works.
Comparing the direct approach (7.96) and the approach suggested in this paper, one can see that the second one reduces the task of inversing non-constant n × n matrix G(y δ(t), t) to the task of solving systems of non-linear ODEs. In some applications numerical integration of ODE systems may be more preferable than matrix inversing. Accurate comparing of this approaches (including numerical computations issues) is the matter of the upcoming studies and also will be published in later works.
7.8 Example
To illustrate the work of the suggested method let’s consider a model of a macroeconomic process, which can be described by a differential game with the dynamics
Here t ∈ [0, T], x 1 is the product, x 2 is the production cost. G(x 1, x 2) is the profit, which is described as
where a 0 = 0.008, a 1 = 0.00019, a 2 = −0.00046 are parameters of the macroeconomic model [1]. The functions u 1(t), u 2(t) are bounded piecewise continuous controls
The control u 1 has the meaning of the scaled coefficient of the production increase speed and u 2 has the meaning of the scaled coefficient of the speed of the production cost changing.
This model has been suggested by Albrecht [1].
We assume that some base trajectories \(x_1^*(t),\ x_2^*(t)\) of system (7.97) have been realized on the time interval t ∈ [0, T] (time is measured in years). This trajectory is supposed to be generated by some admissible controls \(u_1^*(\cdot ),\ u_2^*(\cdot )\). We also assume that we know inaccurate measurements of \(x_1^*(t),\ x_2^*(t)\)—twice continuously differentiable functions \(y_1^\delta (t),\ y_2^\delta (t)\) that fulfill Hypothesis 7.2.
Remark 7.9
To model measurement functions \(y_1^\delta (t)\) and \(y_2^\delta (t)\) real statistics on Ural region’s industry during 1970–1985 [1] have been used. They satisfy Hypothesis 7.2.
We consider the inverse problem described in Sect. 7.5 for dynamics (7.97)–(7.99) and functions \(x_1^*(t),\ x_2^*(t)\), \(u_1^*(\cdot ),\ u_2^*(\cdot )\) and \(y_1^\delta (t),\ y_2^\delta (t)\). We assume in our example that we don’t know the base trajectory and controls, but know the inaccurate measurements \(y_1^\delta (t),\ y_2^\delta (t)\).
The trajectories \(x_1^{\alpha ,\delta }(t),\ x_2^{\alpha ,\delta }(t)\) and controls \(\hat u_1^{\alpha ,\delta }(t),\ \hat u_2^{\alpha ,\delta }(t)\), generating them, were obtained numerically. The results are presented on Figs. 7.1, 7.2, and 7.3. On Figs. 7.1 and 7.2 time interval is reduced for better scaling.
References
Albrecht, E.G.: Metodika postroenija i identifikacii matematicheskih modelej makrojekonomicheskih processov (Methodics of construction and identification of mathematic models of macroeconomic processes). Elektronnyj zhurnal Issledovano v Rossii 5, 54–86 (2002)
Kolmogorov, A.N., Fomin, S.V.: Elementy teorii funktsii i funktsionalnogo analiza. [Elements of the theory of functions and functional analysis] With a supplement, Banach algebras, by V. M. Tikhomirov, 6th edn. Nauka, Moscow, 624 pp. (1989) (in Russian)
Krasovskii, N.N.: Theory of Motion Control. Nauka, Moscow (1968) (in Russian)
Krasovskii, N.N., Subbotin, A.I.: Positional Differential Games. Nauka, Moscow (1974) (in Russian)
Krupennikov, E.A.: Validation of a solution method for the problem of reconstructing the dynamics of a macroeconomic system. Trudy Inst. Mat. Mekh. UrO RAN 21(2), 102–114 (2015)
Kryazhimskii, A.V., Osipov, Y.S.: Modelling of a control in a dynamic system. Eng. Cybern. 21(2), 38–47 (1984)
Osipov, Y.S., Kryazhimskii, A.V.: Inverse Problems for Ordinary Differential Equations: Dynamical Solutions. Gordon and Breach, London (1995)
Subbotina, N.N., Krupennikov, E.A.: Dynamic programming to identification problems. World J. Eng. Technol. 4(3B), 228–234 (2016)
Subbotina, N.N., Kolpakova, E.A., Tokmantsev, T.B., Shagalova, L.G.: The Method of Characteristics for Hamilton–Jacobi–Bellman Equations. Izd. UrO RAN, Yekaterinburg (2013) (in Russian)
Subbotina, N.N., Tokmantsev, T.B., Krupennikov, E.A.: On the Solution of Inverse Problems of Dynamics of Linearly Controlled Systems by the Negative Discrepancy Method. In: Optimal control, Collected Papers. In commemoration of the 105th anniversary of Academician Lev Semenovich Pontryagin, Tr. Mat. Inst. Steklova, vol. 291, MAIK Nauka/Interperiodica, Moscow, pp. 266–275 (2015)
Subbotina, N.N., Tokmantsev, T.B., Krupennikov, E.A.: Dynamic Programming to Reconstruction Problems for a Macroeconomic Model. In: IFIP Advances in Information and Communication Technology. [S.l.], vol. 494, pp. 472–481. Springer, Berlin (2017)
Tikhonov, A.N.: On the stability of inverse problems. Dokl. Acad. Sci. URSS (N.S.) 39, 195–198 (1943)
Tikhonov, A.N., Vasileva, A.B., Sveshnikov, A.G.: Differentsialnye uravneniya [Differential equations]. Fizmalit, Moscow (2005) (in Russian)
Vanko, V.I., Ermoshina, O.V., Kuvirkin G.N.: Variacionnoe ischislenie i optimalnoe upravlenie [Variational calculus and optimal control], 2nd edn. MGU, Moscow (2001) (in Russian)
Acknowledgements
This work was supported by the Russian Foundation for Basic Research (project no. 17-01-00074) and by the Ural Branch of the Russian Academy of Sciences (project no. 18-1-1-10).
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2018 Springer International Publishing AG, part of Springer Nature
About this chapter
Cite this chapter
Krupennikov, E. (2018). On Control Reconstruction Problems for Dynamic Systems Linear in Controls. In: Petrosyan, L., Mazalov, V., Zenkevich, N. (eds) Frontiers of Dynamic Games. Static & Dynamic Game Theory: Foundations & Applications. Birkhäuser, Cham. https://doi.org/10.1007/978-3-319-92988-0_7
Download citation
DOI: https://doi.org/10.1007/978-3-319-92988-0_7
Published:
Publisher Name: Birkhäuser, Cham
Print ISBN: 978-3-319-92987-3
Online ISBN: 978-3-319-92988-0
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)