Iterative Learning Control for Linear Discrete-Time Systems with Randomly Variable Input Trail Length

For linear discrete-time systems with randomly variable input trail length, a proportional(P-) type iterative learning control (ILC) law is proposed. To tackle the randomly variable input trail length, amodified control input at the desirable trail length is introduced in the proposed ILC law. Under the assumption that the initial state fluctuates around the desired initial state with zero mean, the designed ILC scheme can drive the ILC tracking errors to zero at the desirable trail length in expectation sense. The designed ILC algorithm allows the trail length of control input which is different from system state and output at a specific iteration. In addition, the identical initial condition widely used in conventional ILC design is also mitigated. An example manifests the validity of the proposed ILC algorithm.


Introduction
In practice, it is hard to obtain the precise mathematical model of robot system [1][2][3][4], rapid thermal processing [5], flexible system [6][7][8], etc.With simple recursive mode and model-free characteristic, by using the tracking error and control input of the previous iterations, iterative learning control (ILC) is widely used to the dynamical systems with repetitive operations in a fixed time interval [9][10][11].
Hitherto, in order to track the desired trajectory, most existing ILC works require that the trail length is fixed and uniform at each iteration [12][13][14].Nevertheless, in many engineering applications of ILC, the requirement of fixed and uniform trail length might not be fulfilled.The trail length of system output, state, and control input would randomly vary from iteration to iteration due to the complex factors or randomly occurring events.Recently, there are some studies investigating the ILC issues with iteratively variable trail lengths [15][16][17][18][19][20][21][22].By introducing the maximum pass length and adopting the lifted-system framework, Seel et al. [15] proposed the necessary and sufficient conditions of monotonic convergence for linear discrete-time single-input single-output (SISO) systems with varying trail lengths.For linear discrete-time multiple-input multiple-output (MIMO) systems, iteration-average operator and stochastic variable satisfying Bernoulli distribution are involved in ILC design to cope with the varying trail lengths issue [16].It was proved that the system tracking errors can be driven to zero in mathematical expectation sense.Based on the Bernoulli stochastic variable, three varying trail lengths-based ILC schemes were proposed for linear discrete-time systems with vector relative degree [17].In [18], an iteratively moving average operator, which contains the few most recent cycles, and Bernoulli stochastic variable are introduced into ILC scheme for nonlinear continuous-time SISO systems with iteratively varying trail lengths.Other iteratively moving average operator based ILC algorithms for dynamical systems with iteratively varying pass lengths could also be found in [19][20][21].Recently, in [22], the tracking of ILC for discretetime systems with varying trail lengths can be guaranteed in deterministic convergence way by a modified P-type ILC scheme.Among existing varying trail lengths-based ILC studies, it is commonly assumed that the trail lengths of state, control input, and output are identical at a specific iteration.

Complexity
Clearly, more efforts should be made in the varying pass lengths-based ILC algorithms for dynamical systems.
As far as the issue of varying trail lengths is concerned, in most ILC algorithms, it is commonly assumed that the lengths of state, control input, and output are identical at a specific iteration.However, in many practical applications, it is hoped that the controlled system could achieve the control objective with less control efforts.For example, in the speed control of vehicle or robot, when the speed is controlled in a neighborhood of the desired one, the system can operate freely without any control efforts.It implied that the control inputs in a terminal time interval can be removed at one repetitive operation process.This can be represented by a repetitive system with randomly variable control input lengths.Motivated by the above observation, this paper investigates the convergence of varying input lengths-based ILC for linear discrete-time MIMO systems.Based on the assumption that the initial states randomly fluctuate around the desired initial states with zero mean, by applying the proposed P-type ILC law, the ILC tracking errors can be driven to zero in mathematical expectation at the desirable trail length as the iteration number increases.The requirements on identical initial state and trail lengths in the conventional ILC schemes are mitigated in this paper.
The rest of this paper is organized as follows.The ILC issue with randomly variable input trail length is formulated in Section 2. Section 3 presents the P-type ILC law with convergence analysis.An illustrative example is given in Section 4. Section 5 concludes this paper.
Assumption .The iterative initial state   (0) is randomly variable, but its expectation satisfies Remark .Different from the general ILC requirement on identical initial condition that the iterative initial state   (0) =   (0), Assumption 1 is relaxed, where   (0) can be randomly variable with a certain expectation   (0).In this paper, for system (1), the objective of ILC under iteratively varying input lengths and Assumption 1 is to look for an input sequence {  ()},  ∈ {0, 1, . . .,   } such that the ILC tracking can be improved.

ILC Method and Its Convergence
In order to well address the ILC issue of the linear discretetime MIMO system (1) and ( 2) with the iteration-varying input trail lengths, let us denote   (), ( ∈ {0, 1, . . ., }) to be a stochastic variable satisfying Bernoulli distribution and taking binary values 0 and 1.And   () = 1 denotes the event that the control input of system (1) can continue to the time instant  at the -th iteration, which occurs with a probability function of (), (0 < () ≤ 1).  () = 0 denotes the event that the control input of system (1) cannot continue to the time instant  at the -th iteration which occurs with a probability function of 1 − ().
In the following, for the modified system (7) under Assumption 1, a P-type ILC law is used for convergence analysis.
Remark .Regarding the selection of the control gain in the ILC law (8), the convergent condition ( 9) with (10) in Theorem 3 provides us a theoretical guideline.It is noticed that the convergent conditions ( 9) is an inequality, and usually, ILC design does not require the accurate knowledge about the controlled systems.Therefore, based on the estimated values of , , , and (), we can decide the control gain  to satisfy inequality (9).

Illustrative Example
Consider the following discrete-time linear system with randomly variable input trail length, which can be represented as The proposed P-type ILC law (8) is applied.According to the convergent conditions ( 9) and (10) provided in Theorem 3, the control gain for the ILC law is chosen as  = 0.54.The iterative initial states are randomly generated as shown in Figure 1, which satisfy {  (0)} =   (0). Figure 2 presents the performance index   at different iterations.Figure 3 presents the actual tracking situation of the outputs in system (27) and system (28) to the desired trajectory (29) at iterations  = 11 and  = 21 with the P-type ILC law (8).The control inputs with randomly trail length at iterations  = 11 and  = 21 are depicted in Figure 4.It implies that Assumption 1 is very relaxed, where   (0) can randomly fluctuate around   (0) with zero mean.From Figure 2, it is illustrated that the performance index   can be certainly driven to zero as the iteration number  goes to infinity by the proposed P-type ILC law (8).Although the actual input length of system ( 27) is randomly variable with iteration number  as shown in Figure 4 and the initial states fluctuate randomly satisfying {  (0)} =   (0) as shown in Figure 1, a progressively improved tracking performance of the outputs in system (27) and system (28) to the desired trajectory   () has been accomplished for  ∈ {0, 1, . . ., 100} as shown in Figure 3.

Conclusions
For linear discrete-time systems with randomly variable input trail length, where the initial states fluctuate around the desired initial states with zero mean, this paper presents a modified P-type ILC law.Modified control input at the desirable trail length is introduced in ILC designs in order to tackle the randomly variable input trail length.The designed ILC scheme can drive the ILC tracking errors to zero at the desirable trail length in mathematical expectation.The requirements on identical trail lengths at a specific iteration and identical initial condition in previous varying trail length based ILC schemes are mitigated.Future research will extend the varying input trail length based ILC results of this paper to the multiagent systems [23,24] and other robot systems [25].

Figure 2 :
Figure 2: The performance index   at different iterations by using the proposed P-type ILC law (8).

Figure 1
Figure1indicates that the initial condition {  (0)} =   (0) in Assumption 1 is quite different from the identical initial condition   (0) =   (0) of the general ILC schemes.It implies that Assumption 1 is very relaxed, where   (0) can randomly fluctuate around   (0) with zero mean.From Figure2, it is illustrated that the performance index   can be certainly driven to zero as the iteration number  goes to infinity by the proposed P-type ILC law(8).Although the actual input length of system (27) is randomly variable with iteration number  as shown in Figure4and the initial states fluctuate randomly satisfying {  (0)} =   (0) as shown in Figure1, a progressively improved tracking performance of the outputs in system (27) and system (28) to the desired trajectory   () has been accomplished for  ∈ {0, 1, . . ., 100} as shown in Figure3.