OPTIMISING CONTROL ALGORITHMS IN

BIOFEEDBACK-SYSTEMS

First Steps Towards Model Identiﬁcation Adaptive Controllers

Carolin Zschippig, Carsten Rachuy and Kerstin Schill

Group of Cognitive Neuroinformatics, University of Bremen, Enrique-Schmidt-Strae 5, 28359 Bremen, Germany

Keywords:

Adaptive computing, ANFIS, System identiﬁcation, Model Identiﬁcation Adaptive Controllers (MIAC),

Ambient assisted living, Biofeedback, Galvanic skin resistance, Adaptive software.

Abstract:

We explore the potential of model identiﬁcation adaptive controllers (MIAC) within a biofeedback system.

Through the application of an adaptive control algorithm, the system performance could be optimised with

respect to time. In a series of experiments the galvanic skin resistance of test subjects playing a computer

game was recorded. On this data, a system identiﬁcation was performed, utilising an Adaptive-Network-based

Fuzzy Inference System (ANFIS). The results serve as a basis for the development of the adaption laws of the

MIAC and allow conclusions about suitable controllers.

1 INTRODUCTION

Current demographic surveys indicate a shift in the

population structure of Germany, where especially the

relation between younger and elderly people is af-

fected by decreasing birth-rates and increasing life ex-

pectancy. In the year 2060 one third of the population

is expected to be over the age of 65 (Pl

otzsch, 2009).

As a direct result of these demographic changes, the

share of people living in their familiar environments

even in old age will increase accordingly. As aging is

a process which has a negative impact on the physi-

cal and cognitive capabilities of the human body, ac-

tivities of daily living are not only becoming more

difﬁcult to perform but the risk of accidents, med-

ical emergencies and age-related diseases increase.

This poses new challenges for research areas like

telemedicine and ambient assisted living (AAL).

In the former, the user’s physiological conditions

are monitored, with the aim to detect alarming trends,

emergency situations or to provide preventive mea-

sures. In the latter, the aim is to create environments

which support people during their activities of daily

living while simultaneously being capable of recog-

nising emergency situations.

In both cases, physiological signals are either

mandatory or an additional source of information

about the user’s current state.

Preventive measures which ﬁght the onset of fre-

quently encountered health conditions in elderly peo-

ple, like loss of balance, chronic obstructive pul-

monary disease, heart failure or dementia, are often

implemented in systems which actively incorporate

the user through the promotion of physical and mental

exercise (Morris, 2006), (Chiari et al., 2009), (Bosch

et al., 2009), (Oswald, 2007). One approach to de-

signing these systems is using biofeedback, where the

term biofeedback describes a loop consisting of cap-

turing physiological data and the feedback of this data

to the person.

The captured physiological signals are generally

imperceptible to the person on a conscious level, es-

pecially their dynamic properties. Hence the signals

are pre-processed for the feedback and presented to

the person in an easily perceivable manner. Based on

the knowledge about the dynamic physiological sig-

nals, the person is to apply strategies to purposefully

alter his/her physiological state. The most common

styles of presentation are auditory and visual signals

(Rief and Birbaumer, 2006).

In this paper we present a biofeedback system

designed as a game-like computer program which

guides the user into a predeﬁned affective state, so

that the measured data can be utilised as reference val-

ues for later inference processes. Our work is based

on (Rachuy et al., 2011) where we proposed a wear-

able sensor device which is able to measure physio-

logical signals and provide these measurements as in-

put data for successive inference steps about the user’s

affective state.

207

Zschippig C., Rachuy C. and Schill K..

OPTIMISING CONTROL ALGORITHMS IN BIOFEEDBACK-SYSTEMS - First Steps Towards Model Identiﬁcation Adaptive Controllers.

DOI: 10.5220/0003771702070212

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2012), pages 207-212

ISBN: 978-989-8425-88-1

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

2 BIOFEEDBACK-SYSTEM

DESIGN

When designing a biofeedback system, two key fac-

tors are of signiﬁcance: how to represent the captured

data and how to map the dynamics of the data to the

chosen representation.

The decision on how to represent the physiolog-

ical measurements is guided by multiple objectives.

Foremost, the representation should enable the users

to consciously perceive the state of their physiologi-

cal signals and the changes these physiological states

undergo. This can be achieved through a visual rep-

resentation of the data as a factual line plot or numer-

ical value, but the devised requirement of perceptibil-

ity also allows for many more and more abstract di-

agram styles, up to the representation in a game-like

virtual environment (WildDivine, 2010), (SomaticVi-

sion, 2009). Further design objectives are to motivate

the user to engage with the biofeedback system on

a regular basis, despite the repetitiveness of the task.

This deserves particular attention if the biofeedback

system is used as a preventive measure, since the mo-

tivation can not be drawn from the user’s desire to get

relief from a medical condition.

When representing the data in a game-like envi-

ronment the proximity to entertainment technology

facilitates the motivation of continuous engagement.

In these environments a combination of auditory and

visual presentation components can be utilised. Pa-

rameters for the representation of the physiological

states and their dynamics can for example be colour

attributes like chrome or hue, sound attributes like

tempo or volume or the position and movement of ob-

jects. The choices which parameters to use are gener-

ally based on implicit expert knowledge about the aes-

thetics of computer games, formal criteria do hardly

exist (Nacke, 2009).

In addition to the choices of how to represent the

measured values there is the question of how to map

the dynamics of the measured values to changes of the

visual and auditory parameters. The most straight for-

ward and intuitive method is a proportional mapping.

While this approach will lead to a functional biofeed-

back system, it is not necessarily the optimal system

conﬁguration. The application of time-optimising

strategies holds the potential to create a system which

will allow for the user to achieve faster progress in the

alteration of his/her physiological states.

In purely technical systems, the optimisation of

the control strategy is usually done on the basis of a

mathematical description of the system. In a biofeed-

back system, such a description would have to include

a mathematical modelling of the physiological reac-

Figure 1: Block diagram of a MIAC.

tion of the user. Since this reaction can be subject

to the conscious exertion of inﬂuence by the user, it

is highly complex and an analytical approach for a

mathematical modelling is not a feasible strategy. An

alternative lies in the application of model identiﬁca-

tion adaptive controllers (MIAC), schematically de-

picted in ﬁgure 1.

A MIAC can be implemented with very little a-

priori knowledge about the system’s behaviour (Is-

ermann, 1987). Contrary to classical control theory,

the identiﬁcation of the system and the optimisation

of the controller is not done sequentially, but at the

same time. This approach is of advantage if there is

uncertainty of the quality of the identiﬁcation, which

can e.g. be due to stochastic disturbances inhibit-

ing the convergence of the identiﬁcation algorithm.

In the case of biofeedback systems, online identiﬁca-

tion also allows for the incorporation of interpersonal

differences between individuals, since the system be-

haviour does not have to be generalised.

The keyword adaptive describes the property of

the control system to adapt its behaviour according

to the changing properties of the controlled process.

With an adaptive controller a system can be run in a

stable condition, as long as certain preconditions con-

cerning the identiﬁcation method and structure of the

control algorithm are met.

To test and optimise the performance of the cho-

sen identiﬁcation model, adaptation laws and control

structure, the system has to be operated in a closed

loop. In the case of a biofeedback system, this means

it is necessary for a test subject to be present through-

out the optimisation process. To minimise the test-

ing period, we propose an off-line system identiﬁca-

tion, to be able to better estimate applicable controller

structures and adaption laws.

3 SYSTEM IDENTIFICATION

THROUGH ANFIS

The physiological reaction of a person within a

biofeedback loop can be described as dynamic

and non-linear. Therefore, a suitable identiﬁcation

method should be able to approximate dynamic and

non-linear behaviour. Since the system structure and

HEALTHINF 2012 - International Conference on Health Informatics

208

internal states of the system ”human” are not clearly

deﬁnable or measurable, the identiﬁcation is imple-

mented as a so called ”black-box-model” (Schr

oder,

2010). The identiﬁcation is done on the basis of ex-

perimentally gathered data.

In this paper, we describe the implementation of

a Adaptive-Network-based Fuzzy Inference System

(ANFIS), which is a hybrid of fuzzy-logic and neu-

ral networkc, capable of dealing with ill-deﬁned and

uncertain systems (Karray and Silva, 2004).

In an ANFIS, a Sugeno-type fuzzy model is put

into the framework of an adaptive neural network, en-

abling the application of network learning algorithms

(Jang et al., 1997). When implemented as a ﬁrst-order

Sugeno model, the algorithm basically approximates

the dynamic non-linear output function as a weighted

superposition of local linear (time-invariant) models.

The weight assigned to each local model is dependant

on the value of activating parameters, which make up

the input space of the net. Which of the input pa-

rameters are to appear in the local linear models is an

adjustable design parameter. It is therefore possible to

incorporate time-dependency into the activation vari-

ables, but avoid it in the linear models. Also, the input

space can contain parameters which are not control-

lable, but involve an information gain for the validity

area and period of the linear models. If the identiﬁ-

cation of the overall system behaviour as a piecewise

linear, time-invariant function is successful, the online

identiﬁcation within the MIAC can be implemented

as linear time invariant functions. As a consequence,

the control algorithm can also be linear, allowing for

the application of well deﬁned design and optimisa-

tion tools.

4 IMPLEMENTATION

Our biofeedback system was designed as a computer

game. The game was purpose-made for the experi-

ments in order to be able to easily control all rele-

vant variables and integrate measured values in real

time. It was implemented in Java3D. The game genre

was chosen considering the following requirements:

(1) For the adaptive control algorithm it is advanta-

geous if the system is inﬂuenced as little as possi-

ble by disturbance, since the goal of the control algo-

rithm is a good reference transfer behaviour and not

noise rejection. Therefore the game should contain

little elements which inﬂuence the user, but are not

actuating variables. (2) The game should be able to

inﬂuence the user constantly through the same actu-

ating variables. This is facilitated by game mecha-

nisms which provide for the execution of a single task

over a longer time period. (3) The user should not

have the opportunity to deﬁne and pursue meta-goals.

If the user strives to fulﬁl a goal different from the

one predeﬁned by the game design, the control efforts

will be noneffective or even counter-productive. The

game should leave the user little room for manoeu-

vre, meaning it should be of limited complexity. (4)

The game needs to include dynamic elements, whose

variation during the game will not interrupt the game

ﬂow.

Based on these criteria, the game was designed as

a racing game. The user is challenged to steer an

avatar, depicted as a large sphere, keeping it within

a marked track, see ﬁgure 2. The actuating variable

to be controlled by the algorithm is the velocity of the

avatar. Synchronous to a change in velocity the colour

scheme of the game is manipulated. A certain veloc-

ity corresponds to a certain colour, improving the per-

ceptibility of the actuating variable. As an incentive to

perform well on the assigned task, points are awarded

to the player for the position of the avatar. The in-

clusion of this mechanism facilitates a higher immer-

sion of the subjects into the game (Fairclough, 2009).

High immersion ensures that the recorded physiologi-

cal reaction is caused by the game, not environmental

inﬂuences.

5 EXPERIMENT SETUP

In a series of experiments we recorded the physiologi-

cal reaction of test subjects playing the game as input-

output tuples for system identiﬁcation. The physio-

logical parameter examined was the galvanic skin re-

sponse.

The group of test subjects consisted of 14 peo-

ple, 8 male and 6 female. The test subjects volun-

teered for the experiments and did not receive any

reimbursement. All subjects had an academic back-

ground, the mean age of the group was 27,3 years

(Min=20, Max=33, with two subjects refusing to give

precise information). Upon invitation, the users were

informed about health preconditions for participation.

Prior to the experiment and immediately after, the par-

ticipants were asked to ﬁll out a questionnaire regard-

ing personal information and the subjective game ex-

perience.

The users were asked to steer the avatar and try ac-

tively to relax during the game. The instructions were

given in written form. The game was presented on

a 22” LCD-computer monitor, type Samsung Sync-

Master 2233RZ: 120Hz, and speakers placed on the

left and right of the monitor. The user was seated on a

comfortable ofﬁce chair. The experiments were con-

OPTIMISING CONTROL ALGORITHMS IN BIOFEEDBACK-SYSTEMS - First Steps Towards Model Identification

Adaptive Controllers

209

Figure 2: Practise game screen.

ducted in a university laboratory.

The user controlled the movement of the avatar

by acceleration sensors, placed on the back of her/his

dominant hand. Through the rotation of the hand

along the arm axis the avatar was moved horizontally,

from left to right. The velocity of the avatar was the

actuating variable and could not be inﬂuenced by the

user. Prior to the experiment, each user had up to

two minutes to familiarise him-/herself with the game

controls. The game was run in practise mode, with a

constant velocity of 0,6 m/s and the design seen in ﬁg-

ure 2. After the user felt conﬁdent about the controls,

but not later than 2 min, the experiment started. The

variable training time was implemented to accommo-

date the varied experience of the users with computer

games and control through motion.

6 INPUT PARAMETER

SPECIFICATION

Ideally, the experiments should incorporate all possi-

ble states of the actuating variable, so that the system

model covers the entire input space. Due to the fact

that it is a dynamic system, the order of presentation

of the actuation states is of signiﬁcance.

For the ﬁrst set of experiments, we assumed the

model order to be two. For the time-step k, the sys-

tem output y(k) is dependant on the state of the actuat-

ing variable u(k − 1), the previous state of the system

output y(k − 1) and the previous state of the actuat-

ing variable u(k − 2). The physiological reaction to

a stimulus in the galvanic skin response is subject to

latency (Stern et al., 2001). This latency d

has to be

taken into account in the model formulation. With

these assumptions the local linear models become:

y(k) = b

u(k−1−d

)+b

u(k−2−d

)−a

y(k−1)+c

The variation of the actuating variable

velocity+colour during the experiments followed

a predeﬁned schedule. The limits of the range of

values for the actuating variable were determined in

preliminary tests. The continuous range of velocity

values was discertised into 10 equally distanced

values. Out of these values a complete parameter set

of 100 combinations was constructed, consisting of

two consecutive velocities. The tuples were sorted to

form a consistent succession of all possible velocity

changes. During the experiment each velocity was

held constant for 10 seconds, resulting in a total

experiment time of approx. 17 minutes. For each ex-

periment the succession was permuted. The galvanic

skin response was measured on the index and ring

ﬁnger of the subject’s dominant hand. Additionally

to the physiological reaction, we recorded the exerted

acceleration, the position of the avatar, the state of

the actuation variable, the absolute game time and the

awarded points.

We evaluated this data prior to the system iden-

tiﬁcation to examine which parameters should make

up the input space of activating parameters to deter-

mine the validity area and period of the linear models.

For this evaluation we made use of questionnaires and

performed a correlation analysis between the answers

given by the test persons and the recorded data. The

correlation analysis allowed for the following conclu-

sions: (1) The test subjects were under the impres-

sion that their performance in steering the avatar and

keeping it on track improved over the duration of the

experiment. This is in accordance with the recorded

values of the avatar position. Therefore, the absolute

game time is likely to posses an information content

for the validity period of the local models. (2) The

test subjects stated that the position of the avatar (on

track or off track) has had an inﬂuence on their abil-

ity to relax during the game. This supports the hy-

pothesis that the position of the avatar in reference to

the track should be part of the system model’s input

space. (3) The test subjects were inﬂuenced by the

dynamics of the background music. Initially designed

as an element to improve the immersion in the game,

the background music was perceived as having ”fast”

and ”slow” passages. Therefore, the relative volume

of the background music should be assessed regarding

its information content for the validity area. Based on

these conclusions, the input space of the system iden-

tiﬁcation model was set as:







y(k − 1)

u(k − 1 − d

)

u(k − 2 − d

)

onTrack(k − 1 − d

)

music(k − 1 − d

)

time(k − 1 − d

)







(1)

HEALTHINF 2012 - International Conference on Health Informatics

210

7 RESULTS

For the training of the ANFIS a training data matrix

is constructed. Every row of the matrix contains the

input data for the simulation time step k and in the

last column the measured system output y(k + d

) in-

corporating the latency d

. After training, the perfor-

mance of the net is evaluated using a validation data

set, which has the same shape as the training data.

The quality criterion is the average percentage error

with:

Err

∑

measure

− y

net

∑

measure

· 100

In order to obtain a unique solution when opti-

mising the ANFIS through the method of linear least

squares, the number of training data tuples P should

be at least equal to the number of optimised param-

eters m

(Jang et al., 1997). Considering the likeli-

ness of a noise-induced error, it is beneﬁcial to have

P  m

. The number of parameters depends on three

variables:

- the shape of the local models

- the number of net inputs u

- the number of membership functions MBFn

The number of membership functions determines the

granularity of the discretisation of the input range of

each parameter. With the deﬁned shape of the local

models and the input (as eq. (1)), the maximum num-

ber of MBFs is given by:

P  4 ·

∏

i=1

MBFn(u

)

The ﬁrst ANFIS was trained with the structure in

table 1:

Table 1: Structure ANFIS No.1.

MBFn [3, 3, 3, 2, 3, 3]

rule base full

activating function generalised bell

number of training tuples 108928

number of validation tuples 8379

tuples/parameter ratio 56

The average percentage error amounts to

Err

= 16, 53%, with nine-fold cross validation.

The results are summarised in table 2. The change

in the galvanic skin resistance after the presentation

of a stimulus lies at about 5-50% (Stern et al., 2001).

Therefore, it would be desirable to achieve an average

percentage error of less than 5,0 % with every net.

All but one validation produces error values ≤ 10%.

The reason for the signiﬁcantly worse performance

Table 2: Validation results.

ANFIS 1 ANFIS 2

Err

= 56

= 47, 5

V6 3, 76% 39, 4% 1, 6%

V7 10, 28% 275, 4% 1, 7%

V8 5, 22% 20, 9% 1, 2%

V9 110, 02% 247, 2%

V10 4, 75% 21, 2%

V11 2, 78% 20, 0%

V12 1, 64% 5, 0%

V13 4, 86% 22, 1%

V14 5, 47% 86, 2%

Mean 16, 53% 81, 9% 1, 5%

of V9 could not be determined. It could be due to a

faulty attachment of the sensors or the test subject

might a person whose skin conductance does only

very weakly respond to the affective state.

In ﬁgure 3 the best performing and in ﬁgure 4 the

worst reasonably performing validation is depicted.

0 2000 4000 6000 8000

1.5

2.5

3.5

x 10

−3

Data tuples

Output

ANFIS Output

Data

Figure 3: Best performing validation ANFIS 1.

1000 2000 3000 4000 5000 6000 7000 8000

x 10

−3

Data tuples

Output

ANFIS Output

Data

Figure 4: Worst performing validation ANFIS 1.

Taking a closer look at the results of the valida-

tion with the worst but reasonable performance, V7,

it becomes evident that the net’s performance is not

uniform. There are some input value combinations

for which the approximation error is comparatively

larger. This points towards the conclusion that the

partitioning of the input space through the chosen

OPTIMISING CONTROL ALGORITHMS IN BIOFEEDBACK-SYSTEMS - First Steps Towards Model Identification

Adaptive Controllers

211

number of membership functions is not yet ideal and

should be explored further.

To evaluate the inﬂuence of the data tu-

ple/parameter ratio, we trained ANFIS 1 with the

data of 11 out of the 14 experiments, which equates

to a data/parameter ratio of 47,5 (see table 2). The

resulting mean average percentage error amounts to

Err

= 81, 9%, which is signiﬁcantly larger than for

the ﬁrst training. Increasing the number of mem-

bership functions by one for one input would result

in a data/parameter ratio of 41,9. The improvement

gained by a higher input space resolution would be

negated by the low data/parameter ratio, therefore in-

creasing the number of membership functions was

postponed until further data is available.

To further improve the simulation results and to

examine the information content of the chosen acti-

vation variables, ANFIS 2 was trained. Its structure

remained the same as ANFIS 1, except for the input

space, where the parameter time was omitted.

This modiﬁcation improved the net performance

signiﬁcantly, leading to an average percentage error

of only Err

= 1, 5% after three-fold cross valida-

tion. Evaluating the resulting linear functions re-

vealed however that this is achieved by modelling the

output as constant, with the superposition of all linear

functions taking approximately the form y = u

8 CONCLUSIONS

We designed and implemented a biofeedback system

as a game-like computer program to guide the user

into a predeﬁned affective state. As a ﬁrst step to-

wards the development of a model identiﬁcation adap-

tive controller which will optimise the biofeedback

system with respect to time, we conducted a series

of experiments.

Our goal was the approximation of the system as

linear time-invariant models, allowing for the appli-

cation of well deﬁned and efﬁcient optimising tools

when designing the control algorithm and the adap-

tion laws. While being able to generate models with

average percentage errors below 2%, the examined

model structure did not yet reveal a correlation be-

tween the skin conductance and the controllable vari-

ables of the biofeedback game.

However, since the data/parameter ratio showed

to have a high inﬂuence on the net’s performance, we

conclude that for an exhaustive analysis of the inﬂu-

ence of the input partitioning, more data is necessary.

With the higher granularity of the input space, the

generalisation requirement of the linear models is re-

duced. This could result in more accurate models bet-

ter suited to depict the correlation between skin con-

ductance and game parameters.

Further research is suggested to determine

whether an ANFIS can be trained to approximate the

system’s behaviour and to ﬁnd a suitable foundation

to construct a MIAC in which the system is identiﬁed

online as a linear time-invariant model.

REFERENCES

Bosch, S., Marin-Perianu, M., Marin-Perianu, R., Havinga,

P., and Hermens, H. (2009). Keep on moving! activ-

ity monitoring and stimulation using wireless sensor

networks. In EuroSSC’09 Proceedings of the 4th Eu-

ropean conference on Smart sensing and context.

Chiari, L., Farella, E., Rocchi, L., and Benini, L. (2009). A

biofeedback based portable device to support elderly

mobility in the home environment. In World Congress

on Medical Physics and Biomedical Engineering.

Fairclough, S. H. (2009). Fundamentals of physiological

computing. Interacting with Computers, 21:133–145.

Isermann, R. (1987). Digitale Regelsysteme. Springer

Berlin Heidelberg.

Jang, J.-S., Sun, C.-T., and Mizutani, E. (1997). Neuro-

Fuzzy and Soft Computing - A Computational

Approach to Learning and Machine Intelligence.

Prentice-Hall.

Karray, F. O. and Silva, C. W. D. (2004). Soft computing

and intelligent systems design: theory, tools, and ap-

plications. Pearson Education.

Morris, M. (2006). Biofeedback revisited: Dynamic dis-

plays to improve health trajectories. In PERSUASIVE

2006.

Nacke, L. (2009). Affective Ludology: Scientiﬁc Measure-

ment of User Experience in Interactive Entertainment.

PhD thesis, Blekinge Institute of Technology, School

of Computing.

Oswald, W. D. (2007). Pr

avention der demenz – das sima-

projekt. Public Health Forum, 15:28.

otzsch, O. (2009). Bev

olkerung Deutschlands bis 2060

- 12. koordinierte Bev

olkerungsvorausberechnung.

Statistisches Bundesamt, Wiesbaden.

Rachuy, C., Budde, S., and Schill, K. (2011). Unobtrusive

data retrieval for providing individual assistance in aal

environments. In International Conference on Health

Informatics 2011.

Rief, W. and Birbaumer, N., editors (2006). Biofeedback

- Grundlagen, Indikatoren, Kommunikation, praktis-

ches Vorgehen in der Therapie. Schattauer GmbH.

Schr

oder, D. (2010). Intelligente Verfahren: Identiﬁkation

und Regelung nichtlinearer Systeme. Springer Berlin

Heidelberg.

SomaticVision, I. (2009). Somaticvision. abgerufen am

11.11.10.

Stern, R. M., Ray, W. J., and Quigley, K. S. (2001). Psy-

chophysiological Recording. Oxford University Press,

Inc.

WildDivine (2010). Wilddivine. abgerufen am 10.11.10.

HEALTHINF 2012 - International Conference on Health Informatics

212