Predictive Cognitive Modelling of Applications

Sabine Prezenski, Dominik Bruechner and Nele Russwinkel

Cognitive Modelling in Dynamic Human-Machine Systems, Technische Universit

at Berlin,

Marchstraße 23, Berlin, Germany

Keywords:

Cognitive Modelling, Usability, ACT-R, Tool, Usability Criteria, Usability Test.

Abstract:

This paper argues that important usability aspects of mobile applications can be automatically evaluated using

computational cognitive models based on the cognitive architecture ACT-R. A tool incorporating cognitive

models for speciﬁc tasks, users, applications and usability aspects is proposed. Explanations provided by

the tool for usability ﬂaws are based on simulations of cognitive mechanisms. A use-case of the tool is

introduced, which is based on an ACT-R model that simulates how users search and select a speciﬁc target

in a hierarchical android application and predicts efﬁciency and learnability for average users. The model

has been empirically validated in four studies with two different applications. To fully automate the usability

evaluation of the use-case, two basic requirements need to be fulﬁlled. First, the application and the cognitive

model have to be connected. A tool called ACT-Droid acts as an interface between the Android application

and the cognitive model. Second, the models knowledge of the world, which is application speciﬁc, has to be

provided automatically by using an automated user interface analysation approach. Therefore, the open-source

tool AppCrawler was extended to allow the extraction of the required information.

1 INTRODUCTION

More than 2.2 Million applications are available at

Google Play store today in comparison to 1.7 Million

one year ago (AppBrain, 2016). Last year, the num-

ber of applications that were available for Android

increased by over half a million. Every day, over a

thousand new applications appear on the market with

some of them being successful. Upon other aspects,

such as marketing and visibility, very good usability

is a key to successful applications. It appears unfea-

sible that each newly launched app is tested in terms

of usability. Nevertheless, developers ask themselves

if users will like their app and if it is better than those

of their contestants. There are a lot of rivaling appli-

cations and good usability can be the critical factor of

success. Examples of relevant questions addressing

the usability of mobile applications are: How long do

users take to perform a task with a speciﬁc app? Does

user performance increase as users have more practice

with an app? Should an app adapt to the users’ pre-

selection? How should an adaption look like? Will an

update of the app irritate the user? Different methods

(e.g. user testing, expert interviews) provide answers

to these questions. However, these methods are both

time and money consuming, and the results often do

not provide a clear explanation of the ﬁndings. Fur-

thermore, changes in an app often require a new eval-

uation. We are currently developing a fully automated

approach for Android applications that will automati-

cally predict important usability aspects (such as efﬁ-

ciency) based on cognitive models that simulate cog-

nitive processes of users. Our approach will further-

more offer insight into the cognitive reasons behind

usability ﬂaws.

We want to present a prototype aimed on one spe-

ciﬁc use-case, namely the automated usability evalu-

ation of a hierarchical style Android application for

an average user (Prezenski and Russwinkel, 2016).

Therefore, it is required to extend an existing cog-

nitive model (Prezenski and Russwinkel, 2016) to

be applicable to a multitude of hierarchical list style

applications without any prior modiﬁcation of the

model. Thus, any model aspects that are speciﬁc for

the application (such as relevant semantic knowledge)

should to be automatically extracted from the user in-

terface of the mobile application.

In this paper ﬁrst a theoretical introduction about

usability and cognitive models in HCI is given. Sec-

ondly, a general automated usability evaluation ap-

proach based on ACT-R cognitive models is intro-

duced. This is followed by the deﬁnition of a more

speciﬁc use-case for hierarchical list style applica-

tions to base a working prototype on. Requirements

Prezenski S., Bruechner D. and Russwinkel N.

Predictive Cognitive Modelling of Applications.

DOI: 10.5220/0006273301650171

In Proceedings of the 12th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2017), pages 165-171

ISBN: 978-989-758-229-5

165

on the approach are then deﬁned to make it applicable

on our use-case. Finally, future steps are elaborated

followed by a discussion.

2 THEORY

2.1 Usability

The term usability deﬁnes how efﬁcient, effective,

and satisfying the use of a technical system is (ISO-

9241-11). The PACMAD (People at the Centre of

Mobile Application Development) complements this

usability deﬁnition with aspect important for mobile

usability, namely learnability, memorability, users,

task, context, and mental load (Harrison et al., 2013).

Note that the usability of mobile applications should

not be seen as a smaller desktop, but different aspects

need to be considered.

While applications exist for almost every imagin-

able task, from selecting a recipe to planning a va-

cation, a single app mostly supports only one or two

main tasks (e.g. messenger applications are mainly

used for reading and sending messages or recipe ap-

plications for searching for recipes). And although

on the surface applications appear incredible divers

(e.g. different colors, font, framing of images), the

core principles (e.g. menu-structure, organization) are

similar for a large amount of applications. In our view

these underlying core structures have a strong inﬂu-

ence on many usability criteria. This similarity of

structures in a wide range of applications and the fo-

cus on one or two core functions for most applications

is why we propose that usability of applications can

be analyzed using computational cognitive models.

We aim to develop cognitive models for all aspects of

PACMAD, but will begin developing models for the

usability aspects of efﬁciency and effectiveness and

learnability. In (Nielsen, 1994) a set of nine heuristics

for different usability aspects were presented which

can be applied during the application design process.

In the past these were not automated testable. For

some of those heuristics for good usability cognitive

models could be developed as well.

2.2 Cognitive Models

In our understanding, computational models are cog-

nitive models if they can simulate human behaviour

by simulating ongoing cognitive processes and mech-

anisms both on a symbolic and sub-symbolic level.

We want to simulate the process of humans in order

to understand and predict their behaviour.

2.2.1 ACT-R

Our cognitive modelling approach uses the cognitive

architecture ACT-R (Adaptive Control of Thought

Rational) (Anderson, 2007). In general terms, cog-

nitive architectures are computational platforms that

allow the creation of models on the basis of how hu-

mans process information. Cognitive architectures

constrain models in a way that only cognitively fea-

sible models are possible. More than simply repro-

ducing the behavioural output, cognitive architectures

reveal underlying cognitive mechanisms. ACT-R is

the most widely used cognitive architecture and it has

been applied in many domains from air trafﬁc control

(Raufaste, 2006) to mathematical tutors (Ritter et al.,

2007). It is open source and used by an active com-

munity of researchers. The ACT-R approach aims to

achieve an uniﬁed theory of cognition in the future,

and is the most successful computational approach of

modelling human cognition on a symbolic level that

we are aware of.

The ACT-R architecture consists of different mod-

ules (representing the modular structure of the human

processing) and buffers (the communication inter-

faces of these modules). See Figure 1 for an overview

of the most important buffers (arrows) and modules

(boxes) of ACT-R. For example, there is a motor mod-

ule for motor processes such as key-presses and a vi-

sual module (for visual processing). ACT-R also has

a working memory module (imaginal module), which

is important for learning and combining information

and a declarative memory module where knowledge

(chunks) is stored and retrieved from.

Writing an ACT-R model normally requires the

modeler to specify the models’ fact knowledge man-

ually (declarative knowledge or chunks) and its rules

(production rules, procedural knowledge). These pro-

duction rules are the core part of the model. They

consist of an if and a then part. The if part refers to

the state and content of the buffers (e.g. if they are

used and if they hold speciﬁc chunks) and the then

part modiﬁes the content and state of the buffers.

Next to the productions rules, the processing of

information is governed by numerous sub symbolic

processes, which have been added to the architec-

ture only after empirical studies have conﬁrmed them.

For example, the retrieval of chunks from declarative

memory is determined by the sub-symbolic activation

value of the chunk, which depends on when the chunk

was last used.

2.2.2 Cognitive Models on HCI Aspects

A number of ACT-R models concern menu search or

mobile interaction: (Byrne et al., 1999) demonstrated

HUCAPP 2017 - International Conference on Human Computer Interaction Theory and Applications

166

Figure 1: Here a selection of the modules (boxes) and

buffers (arrows) of the ACT-R architecture is represented.

Note that some modules have two buffers (e.g. the visual

module).

that ACT-R models are capable of humanlike menu

search in a controlled study. In this study, perfor-

mance in a random menu selection task on a computer

was compared with eye tracking and behavioural data.

(Das and Stuerzlinger, 2007) modelled the learning

of mobile phone usage for elderly novice users. The

increase in performance in their model was real-

ized through remembering locations of keys. (Gal-

lagher and Byrne, 2015) investigated password en-

try on smartphones and offered model-based explana-

tions on how requirements for complex passwords on

smartphones should be designed to meet both users’

abilities and safety requirements. These studies indi-

cate that expert and novice behaviour, menu search,

as well as learning behaviour can be simulated using

cognitive architectures.

2.2.3 Modelling Approaches of Usability

The modelling approaches presented above show only

limited usability prediction based on existing cog-

nitive architectures. The most successful approach

is CogTool (John et al., 2004) and its’ successor

CogTool Explorer (Teo and John, 2008). These ap-

proaches can predict expert and novice search be-

haviour on websites. Some aspects of the approaches

are based on ACT-R, but they do not use the full

power of ACT-R due to a lack in memory and learning

mechanisms. Thus usability aspects based on learning

or related aspects, such as repeated usage, cannot be

addressed by the CogTool approaches. Furthermore,

much manual work (e.g. preselecting ideal paths) is

required when working with CogTool.

Other modelling approaches of user behaviour are

GOMS-based cognitive models. GOMS models are

often used in HCI since their implementation is ef-

fortless and simple. Modelers specify Goals, Opera-

tors, Methods and Selection Rules to simulate user be-

haviour. Many GOMS models provide an acceptable

prediction on the time that skilled users need for a pre-

deﬁned task. The straightforward approach of GOMS

models makes them convenient to use; task are di-

vided into subsets of operators with assigned time

value. Although these models are declared ”cogni-

tive” they mainly consist of different motor and visual

operators. A single mental operator (with a speciﬁc

time value) represents the entire cognitive process of

”thinking”, which is oversimplifying for many evalu-

ation questions. GOMS models are a great approach

whenever the objective is to ﬁnd out how much time a

skilled user will need for a task. Other (non-cognitive)

usability modelling tools exist as well, in that they use

different computational algorithms to analyze aspects

such as font-size, color, contrast (Amalﬁtano et al.,

2012; Choi et al., 2013).

Two major challenges have prevented modelers

from evaluating usability of mobile applications with

cognitive architectures and are the reasons why tools

with diminished explanatory power have been used

instead. The ﬁrst challenge is the necessity to (re)-

construct an interface the model can interact with.

We have developed a new tool, ACT-Droid (D

orr

et al., 2016), that directly connects Android applica-

tions with an ACT-R environment, making mock-up

creations or translations obsolete. The second chal-

lenge is the high amount of expertise that is currently

required to construct cognitive models with cognitive

architectures. The more different the applications are

in terms of general interaction mode, the more effort

it takes for the modeler to transfer a model to another

app even if the cognitive mechanisms in the model

(e.g. learning mechanism) are the same. Therefore,

we are focusing the modelling work on transferable

cognitive models. These can be reused for other sim-

ilar applications without effort. Such a model exists

for hierarchical list style applications. It has success-

fully predicted empirical user behaviour (such as re-

peated application usage) for two different applica-

tions. The studies and results are described in more

detail in (Prezenski and Russwinkel, 2016) and are

summarized in the following: The two applications

were a shopping-list application, which allows users

to select items out of different stores and categories

and a real-estate application from which users can se-

lect search criteria (e.g. search for an apartment with

60m). Four empirical user studies, two with each ap-

plication, were conducted. The users were required

to repeatedly search for seven criteria with the same

version of the application twice, and then the applica-

tion was modiﬁed either due to an update (shopping-

list application) or due to adaptation to prior selected

search criteria (real estate application). Two different

versions of both applications existed – thus, in each

of the four studies, the participants started with a dif-

Predictive Cognitive Modelling of Applications

167

ferent application. The ACT-R model predicts mean

overall item selection time of users for ﬁrst and sec-

ond time search. Furthermore it predicts search be-

haviour after updates accurately in terms of correla-

tion and mean standard deviation (see Figure 2, orig-

inally appeared in (Prezenski and Russwinkel, 2016,

p. 205).

selection time (sec)

house apartment

participants

model

r²=1.000

RSME=0.011

n=12

selection time (sec)

apartment house

participants

model

r²=0.888

RSME=1.660

n=10

selection time (sec)

shopping 3-4

participants

model

r²=0.799

RSME=1.489

n=17

selection time (sec)

shopping 4-3

participants

model

r²=0.999

RSME=0.843

n=12

run

house

apartment

4 layers

3 layers 3 layers

3 layers

update

adaptation

Figure 2: Mean target selection time for the four different

studies for the modelled and empirical data (Prezenski and

Russwinkel, 2016).

3 PROPOSED USABILITY

EVALUATION APPROACH FOR

MOBILE APPLICATIONS

3.1 Overall Aim

Our longterm goal is to develop a toolset which can

automatically access important usability aspects of

mobile Android applications as deﬁned in PACMAD.

In the development phase, features will be added sub-

sequently and evaluated empirically.

Our approach is aimed at helping developers eval-

uate design alternatives in terms of cognitive aspects

of usability. The toolset will output numerical val-

ues for different usability criteria such as efﬁciency

(mean time on task) and effectiveness (proportion of

successful attempts on whole attempts). Compara-

tive values will be incorporated in the future as well.

See Table 1 for an overview on the toolsets’ proposed

functionality. Testing an application will require the

tool-users to pre-select certain application properties

as shown in Table 1.

Table 1: Proposed functionality.

Application type text-style

menu type - hierarchical

icon-style

...

Usability criteria efﬁciency

effectiveness

learnability

cognitive load

satisfaction

...

Type of task search & select speciﬁc target

search & select unspeciﬁc target

...

Type of user average user

elderly user

initial use

second time use

context

...

The level of detail of the output will also depend

on the developers previous selections. The output will

consist of numerical values (such as mean and stan-

dard deviations, a comparative value in relation to

other similar applications) and a more cognitive out-

put, revealing potential cognitive causes of usability

ﬂaws (e.g. retrieval failures for uncommon words).

For different applications (e.g. icon based vs. text

based), different tasks (e.g. search and selection of a

speciﬁc targets, search for an unspeciﬁc target), and

different usability concerns (e.g. efﬁciency, learnabil-

ity), the cognitive modelling community has come up

HUCAPP 2017 - International Conference on Human Computer Interaction Theory and Applications

168

with some solutions; however, much work is still re-

quired.

To move towards a fully automated usability eval-

uation approach based on cognitive models, we want

to present a prototype for a certain use-case in this

paper. Our use-case is the automated usability eval-

uation of a hierarchical list style Android application

(cf. (Prezenski and Russwinkel, 2016) model for an

average user). The model makes predictions for initial

and repeated use and can evaluate the usability crite-

ria efﬁciency (average time on task) and learnability

(comparing ﬁrst and further time on task). The model

supports the selection of speciﬁc targets as a task.

3.2 Current State

3.2.1 ACT-Droid

ACT-R is written in the programming language LISP,

and mobile applications are mostly developed for ei-

ther iOS or Android. Enabling interaction between

ACT-R models and mobile applications is crucial.

To allow the model interaction for Android, ACT-

Droid was developed (D

orr et al., 2016). ACT-Droid

uses TCP/IP sockets to directly link ACT-R models

with Android applications. Currently, we are working

on integrating Touch-Commands for ACT-R (Greene

et al., 2013) into ACT-Droid. Touch-Commands, such

as swipe and peck, have been implemented in ACT-R

with the ACT-Touch approach.

Figure 3: Overview of the Shopping-List Application.

3.2.2 Android Demo Application

A model to automatically evaluate cognitive aspects

of usability in mobile applications has been devel-

oped (Prezenski and Russwinkel, 2016). We tested

the model with a shopping-list demo application (see

Figure 3). It allows searching for items which are or-

ganized in categories and subcategories. E.g. if a per-

son (or the model) wants to select alcohol-free beer

for their shopping list, they must ﬁrst select shops,

then bottleshop and then beer before they can select

the target alcohol-free beer. Each subcategory is on a

different page of the app.

3.2.3 ACT-R Model

When the model searches for a target the ﬁrst time

(initial use), it does not know which subcategory it

has to select in order to ﬁnd its targets. Thus, ev-

ery item of each category is read and for each item

the model uses its knowledge of the world to check if

the target can be found under the current category. In

other words, if there is an association between the tar-

get and the current item. For example, the model sees

the word ”vegetables”. It will check its knowledge of

the world if there is connection between ”vegetables”

and its target ”alcohol-free beer”. If it can not ﬁnd a

connection in its’ knowledge of the world, it will read

the next item ”beer” and search for a connection in its’

knowledge of the world. If it ﬁnds a connection, the

model will select the category, remembering the po-

sition of the category and build up a path (in working

memory) containing the categories leading to the tar-

get. Thus, it builds up experience and can later (e.g.

when it is looking for the same target again) use its

experience to navigate the target using the paths and

position chunks stored in its declarative memory. Sub

symbolic mechanisms, such as activation, inﬂuence

if the chunks (paths, position and associations from

world knowledge) can be retrieved.

Currently, the knowledge of the world chunks need

to be added by hand for every new app which needs

manual effort.

3.3 Towards a Fully Automated

Interface Evaluation Approach

In order to create a working prototype of our auto-

mated approach, it is required to extend the current

ACT-R model to be applicable to a multitude of hier-

archical list style applications without prior modiﬁca-

tion. Furthermore, we need to supply the knowledge

of the world to the ACT-R model by automatically

extracting it from the user interface of the mobile ap-

plication. The data ﬂow of our approach is presented

in Figure 4. ACT-Droid acts as an interface between

!"#$%!"#$&'()*

!"#$%&

'()$*+,-$

.'/0

122&

'()$*+,-$

3141

!+*'()*

!,,-)./0)(+

0*"#5-)6"(

75%$8

9(":%$#;$<

"+<)=$<>"*%#

123'$4+03'5/.36!+/-72)2

Figure 4: Data ﬂow.

Predictive Cognitive Modelling of Applications

169

the Android application and ACT-R. Additionally an

UI analysis is executed beforehand (see 3.3.2).

3.3.1 Extension of the Current Model

As a next step we want to demonstrate that automati-

cally evaluating cognitive aspects of usability in mo-

bile applications is feasible using the (Prezenski and

Russwinkel, 2016) model. We will extend the model,

so that it will be applicable for all hierarchical list

style applications without any manual modiﬁcation.

Thus, we can automatically predict efﬁciency, effec-

tiveness, and learnability for ﬁrst and further time use

of these applications. In order to understand which

of the models’ aspects need to be modiﬁed, the main

model mechanism and an exemplary app which the

model processes will be explained brieﬂy. For a more

detailed description of the model, see (Prezenski and

Russwinkel, 2016). The model is for menu-based

hierarchical applications that are text-based. It sim-

ulates search and speciﬁc target selection and can

model both initial and repeated use.

3.3.2 Automated UI Analyzation to Create

World Knowledge

Currently, the knowledge of the world chunks need to

be added by hand for every new app. We want to

automate this process using a program that automat-

ically enables the creation of new chunks in ACT-R

declarative memory module. The chunks will be the

connection nodes between all possible targets in an

app and the categories and subcategories leading to

the targets. The chunks will always contain only two

elements (e.g. the target and a category).

To acquire this knowledge of the world chunks

automatically, it is required to analyse the complete

user interface of the mobile app to evaluate before-

hand. For our ﬁrst use-case, the analyzation of an

android app, an OpenSource approach to crawl user

interfaces called AppCrawler, was extended. This ap-

plication is an automatic UI testing tool for Android

based on UIAutomator, a UI testing framework suit-

able for cross-app functional UI testing across system

and installed Android applications.

AppCrawler uses a Depth-ﬁrst search algorithm

(Cormen, 2009) to traverse through an apps interface.

DFS is an algorithm for traversing or searching tree or

graph data structures (Cormen, 2009). By using the

DFS for our use-case, every screen of the application

is traversed.

The existing implementation needed to be ex-

tended to extract the text value of each clickable menu

item. Additionally, it was necessary to extend the

AppCrawler to store all menu-item to target relations

which are then used to create the knowledge of the

world for the ACT-R model. Targets in this case are

the leaves of the search tree, e.g. alcohol-free beer,

menu-item to target relations would be bottle-shop -

alcohol-free beer and beer - alcohol-free beer. The

automatically extracted information can then be sup-

plied as knowledge of the world chunks to the ACT-R

model.

3.3.3 Pre-activation of Targets

In a further step, the pre-activation of these targets

will be weighted depending on how often they appear

in the same sentence in written language databases

(such as Wikipedia). Such an approach is used in

combination with ACT-R and described in more de-

tail in CogTool Explorer (Teo and John, 2008).

4 DISCUSSION

Some questions remain open and more work is re-

quired until a developers’ version of the tool can be

launched. For many aspects of the tool, cognitive

models still need to be developed and evaluated.

Currently, we are exploring possibilities to model

elderly users with ACT-R and incorporate such mech-

anism into the tool.

We would also like to extend the cognitive models

for different cultural backgrounds (e.g. reading direc-

tion in arabic).

Furthermore, the UI analyzation approach needs

to be extended to cover different kinds of app layouts

(e.g. text-style, icon-style) as well as different app

development platforms (e.g. iOS).

As a next step, we want to apply the modelling

approach of the use-case to a common used applica-

tion and compare the model results with user data. If

this is successful, this is a strong proof of concept that

ACT-R-based cognitive modelling is indeed useful for

usability testing.

We are often asked why we attempt to model with

ACT-R instead of using machine learning approaches

to simulate user behaviour. With enough training and

data, neural network approaches (e.g. machine learn-

ing algorithms) can probably predict user behaviour

as well; however they cannot offer any insight into

why these results are obtained. The cognitive pro-

cesses of users leading to the outcome are a black

box. Thus, no advice on cognitive aspects to design-

ers on how to change the applications can be given

by such an approach in comparison to an ACT-R ap-

proach. Given that the use-case is fully automated

(see 3.3.2) and pre-activation of the knowledge of

HUCAPP 2017 - International Conference on Human Computer Interaction Theory and Applications

170

the world chunks is available (see 3.3.3), an ACT-R

model can help with ﬁnding problems related to cog-

nitive aspects. Failures of a model related to a missing

knowledge of the world chunk for a target (e.g. the

model is searching for alcohol-free-beer, it looks at

bottle shop but can’t retrieve a knowledge of the world

chunk linking bottle shop to to alcohol-free beer) can

be interpreted to solve those problems. For example,

a retrieval failure of the knowledge of the world chunk

can occur because it had too low of an activation value

and this should be interpreted as a bad labeling choice

(rename bottle shop).

Furthermore, our approach (other than machine

learning approaches) does not require a large amount

of data to work. Once the ACT-R model has been

tested for smaller empirical sample sizes (e.g. 20 par-

ticipants) it is possible to estimate reliable usability

criteria if it predicts the main usability criteria accord-

ingly. If the ﬁt to the empirical data is satisfying, the

model should be tested with another similar app, and

if it is successful again, it will be incorporated in the

tool.

5 LIMITATIONS

ACT-R’s level of detail on predicting visual processes

is not detailed enough to simulate how complex visual

information is processed and perceived, as is used in

map or game applications. Thus, deciphering usabil-

ity effects of such applications is out of the scope of

our tool. We are looking for cooperations with other

usability approaches that focus more on analyzing vi-

sual processing.

REFERENCES

Amalﬁtano, D., Fasolino, A. R., Tramontana, P.,

De Carmine, S., and Memon, A. M. (2012). Using gui

ripping for automated testing of android applications.

In Proceedings of the 27th IEEE/ACM International

Conference on Automated Software Engineering, ASE

2012, pages 258–261, New York, NY, USA. ACM.

Anderson, J. R. (2007). How can the human mind occur in

the physical universe? Oxford University Press, USA.

AppBrain (2016). Number of available android appli-

cations. https://www.appbrain.com/stats/number-of-

android-apps. Retrieved 2016-01.

Byrne, M. D., Anderson, J. R., Douglass, S., and Matessa,

M. (1999). Eye tracking the visual search of click-

down menus. In Proceedings of the SIGCHI Confer-

ence on Human Factors in Computing Systems, CHI

’99, pages 402–409, New York, NY, USA. ACM.

Choi, W., Necula, G., and Sen, K. (2013). Guided gui test-

ing of android apps with minimal restart and approxi-

mate learning. SIGPLAN Not., 48(10):623–640.

Cormen, T. H. (2009). Introduction to algorithms, chapter

22.3: Depth-ﬁrst search, pages 540–549. MIT press.

Das, A. and Stuerzlinger, W. (2007). A cognitive simulation

model for novice text entry on cell phone keypads.

In Proceedings of the 14th European Conference on

Cognitive Ergonomics: Invent! Explore!, ECCE ’07,

pages 141–147, New York, NY, USA. ACM.

orr, L.-M., Russwinkel, N., and Prezenski, S. (2016). Act-

droid: Act-r interacting with android applications. In

Reiter, D. and Ritter, F. E., editors, Proceedings of the

14th International Conference on Cognitive Modeling

(ICCM 2016), pages 225–227, University Park, PA:

Penn State.

Gallagher, M. A. and Byrne, M. D. (2015). Modeling pass-

word entry on a mobile device. In Proceedings of the

2015 International Conference on Cognitive Model-

ing.

Greene, K. K., Tamborello, F. P., and Micheals, R. J. (2013).

Computational cognitive modeling of touch and ges-

ture on mobile multitouch devices: Applications and

challenges for existing theory. In International Con-

ference on Human-Computer Interaction, pages 449–

455. Springer.

Harrison, R., Flood, D., and Duce, D. (2013). Usability

of mobile applications: literature review and rationale

for a new usability model. Journal of Interaction Sci-

ence, 1(1):1.

John, B. E., Prevas, K., Salvucci, D. D., and Koedinger,

K. (2004). Predictive human performance modeling

made easy. In Proceedings of the SIGCHI conference

on Human factors in computing systems, pages 455–

462. ACM.

Nielsen, J. (1994). Enhancing the explanatory power of

usability heuristics. In Proceedings of the SIGCHI

Conference on Human Factors in Computing Systems,

CHI ’94, pages 152–158, New York, NY, USA. ACM.

Prezenski, S. and Russwinkel, N. (2016). Towards a gen-

eral model of repeated app usage. In Reiter, D. and

Ritter, F. E., editors, Proceedings of the 14th Inter-

national Conference on Cognitive Modeling (ICCM

2016), pages 201–207, University Park, PA: Penn

State.

Raufaste, E. (2006). Air trafﬁc control in act-r: A compu-

tational model of conﬂict detection between planes.

In Proceedings of the International Conference on

Human-Computer Interaction in Aeronautics (HCI-

Aero’06), pages 258–259.

Ritter, S., Anderson, J. R., Koedinger, K. R., and Corbett,

A. (2007). Cognitive tutor: Applied research in math-

ematics education. Psychonomic bulletin & re-

view, 14(2):249–255.

Teo, L. and John, B. E. (2008). Cogtool-explorer: to-

wards a tool for predicting user interaction. In CHI’08

Extended Abstracts on Human Factors in Computing

Systems, pages 2793–2798. ACM.

Predictive Cognitive Modelling of Applications

171