**4.2 Robust (H ∞ ) Control**

**4.2.4 Robust Control With Sensitivity Analysis**

In this subsection I will present the results of robust control which was designed using
sensitivity analysis. These results are not only my own results. In the current subsection
I will describe only partial results which are closely connected to my *Thesis group 1,*
not the whole controller design methods and full simulation results will be discussed.

The aim is to investigate how the designed controller reacts to parametric changes, viz.

robustness of the controller.

The current robust control was designed for the tumor growth model under angiogenic
inhibition described in (3.4-3.6), and it was carried out with sensitivity analysis in order
to determine the uncertainty weighting matrix, and to investigate the effect of parametric
perturbation on the closed-loop system. The tumor growth model was linearized at the
*x*10= 100 mm^{3} operation point similarly as in Subsection 4.2.1. Results were published
in A. Szeles, J. S´api, et al.2012; A. Szeles, D. A. Drexler, et al.2014; L Kov´acs, A Szeles,
et al. 2014.

**Sensitivity Analysis**

Starting from the formal definition of the multiplicative uncertainty, parametric sensitivity
analysis was performed on the nonlinear model to determine*W** _{unc}*. The idea was partially
adapted and modified from (L Kov´acs, Kulcs´ar, et al. 2011; Liu and Zeng 2012) in order
to incorporate uncertain parameters into the design process. Ranges are associated to
these selected parameters. By taking every single extremal combination of the parameters,
linearization is performed. Finally, the frequency content of the perturbed and linearized
model is compared and relative difference is computed. Instead of using the extremal
values, a gridding technique is proposed. Consequently, we consider the combination of
the selected parameters in a multiplicative manner.

A ±5% variability of the Lewis lung carcinoma parameters and a ±10% variability of
the vascular inactivation rate was assumed (Hahnfeldt et al. 1999). For the Lewis lung
carcinoma parameters, factors*f*_{1}, *f*_{2} and *f*_{3} were chosen from a (+5%, +2.5%, 0,−2.5%,

−5%) grid, and for the vascular inactivation rate, factor *f*4 was taken from a (+10%,
+5%,−5%, −10%) grid. The perturbed nonlinear model is:

0.0010 0.01 0.1 1 10 100 0.2

0.4 0.6 0.8 1 1.2

1.4 Sensitivity analysis

Multiplicative magnitude error

Frequency (rad/day)

Figure 4.13: Relative modeling error functions (perturbed system compared to the nom-inal model) in frequency domain and uncertainty upper bound (dashed line).

*x*˙_{1}(t) = −λ_{1}*x*_{1}(t) log *x*_{1}(t)
*x*_{2}(t)

!

(4.66)
*x*˙_{2}(t) = *bx*_{1}(t)−*dx*_{1}(t)^{2/3}*x*_{2}(t)−*ex*_{2}(t)u(t) (4.67)

*y* = *x*_{1}*,* (4.68)

where *λ*_{1} = (1 +*f*_{1})*λ*_{1}, *b* = (1 +*f*_{2})*b,* *d* = (1 +*f*_{3})*d* and *e* = (1 +*f*_{4})*e. For each*
possible combination, the nonlinear model was linearized at the*x*10= 100 mm^{3} operation
point, and the obtained linear model was used to determine parametric sensitivity by
determining*supW** _{rel}* of the relative uncertainty relation:

*W**rel*(ω) =

*G** _{p}*(ω)−

*G*

*(ω)*

_{n}*G*

*n*(ω)

*,* (4.69)

where *G** _{p}* stands for the perturbed model and

*G*

*for the nominal model. The frequency*

_{n}0 5 10 15 20 25 30 35 40

Figure 4.14: Characteristics of tumor regression and control input in case of different perturbation scenarios – parameters change between the 5th and 10th day (blue), the 10th and 15th day (red) and the 15th and 20th day (green), each

model parameter is perturbed independently with a variability of±25%.

domain of interest was*ω*∈[0.001 100] rad/day. The determined parametric sensitivity
was upper bounded (Figure4.13):

*supW**rel*= 0.47*s*+ 8

*s*+ 2*.* (4.70)

The results of the sensitivity analysis show that in lower frequency domain, *ω* ∈
[0.001 0.1] rad/day, the model is less sensitive for parameter perturbation than in higher
frequency domain, *ω* ∈ [0.01 100] rad/day. In the low frequency domain which is
characteristic for tumor growth dynamics, the uncertainty upper bound allows 125%

deviation in the gain of the transfer function relative to the nominal transfer function (Figure4.13) instead of the 65% deviation resulted from the perturbation of the parameters.

This means that significantly larger variance of the parameters is allowed in the low frequency domain, regardless of their dynamical characteristics.

The obtained uncertainty weighting function *W* should work as a high pass filter to

reduce disturbance at low frequency, and to avoid strong restrictions at high frequency:

*W**unc*= 0.01*s*+ 2

*s*+ 8*.* (4.71)

Sensor noise, as a wide-band signal, can be modeled with a constant value. During the
design process,*W**n*anticipates 5% measurement noise for volume measurements. This is
in accordance with the measurement noise used in (Hahnfeldt et al. 1999).

*W**n*= 0.05. (4.72)

The control input is penalized by the weighting function *W** _{u}*, wchich was chosen to

*W** _{u}* = 0.01. (4.73)

The zero of the weighting function was chosen based on the uncertainty weighting function to form the desired ”cone” shape in frequency domain. The amplification was set to compensate the amplification of the model-matching function and to minimize oscillation in constant reference signal tracking:

*W** _{perf}* = 6.5·10

^{−7}

*s*+ 2

*s*+ 8*.* (4.74)

**Effect of Parametric Perturbation on the Closed-Loop System**

Effect of parametric perturbation on the closed-loop system was investigated. In this
case, there is no measurement noise, and the measurements are taken continuously. The
parameters are perturbed independently (b, *d,λ*_{1}, *e) with a variability of* ±25% in three
different time intervals:

• between the 5th and 10th day,

• between the 10th and 15th day,

• between the 15th and 20th day.

Parameter perturbation does not affect tumor regression before the 5th day because of the applied saturation. After the 20th day, steady state is achieved and the tumor volume is nearly minimal; thus, perturbations do not change significantly the performance of the controller.

If the condition of the patient changes after the 10th day of the therapy, the speed of tumor regression does not change remarkably, total inhibitor inlet varies between 815.8

5−10 10−15 15−20 700

750 800 850 900 950

1000 Total inhibitor inlet in case of transient behavior

Total inhibitor inlet (mg/kg)

Days of transient

Figure 4.15: The total inhibitor inlet in case of different perturbation scenarios – param-eters change between the 5th and 10th day (blue), the 10th and 15th day (red) and the 15th and 20th day (green), each model parameter is perturbed

independently with a variability of±25%.

mg/kg and 916.4 mg/kg (Figure4.15). Parameter perturbation between the 5th and 10th day can cause both deterioration and amelioration in terms of speed of tumor regression, daily and total inhibitor inlet (Figure 4.14). The total inhibitor inlet varies between 742.9 mg/kg and 964.9 mg/kg (Figure 4.15). In the figure each cross represents a total inhibitor inlet value for a perturbed parameter, e.g. one cross means the total inhibitor inlet if the tumor growth rate (λ1) is perturbed with +25%. One can see from the figure that the effect of the perturbation decreases as the treatment time increases.

These simulations demonstrated that the designed controller reacts to parametric changes very pliantly, as expected from robust control methodology.