BenMartine.doc

Soft Computing and User Modeling

T. P. Martin1 and B. Azvine2
1
University of Bristol, Bristol, BS8 1TR, UK
Trevor.Martin@bristol.ac.uk
2
BTexact Technologies, Adastral Park, Ipswich, IP5 3RE, UK
Ben.Azvine@bt.com

Abstract. The next generation of consumer goods, including computers, will be
much more sophisticated in order to cope with a less technologically literate user
base. A user model is an essential component for “user friendliness”, enabling the
behavior of a system to be tailored to the needs of a particular user. Simple user
profiles already personalise many software products and consumer goods such as
digital TV recorders and mobile phones. A user model should be easy to initialise,
and it must adapt in the light of interaction with the user. In many cases, a large
amount of training data is needed to generate a user model, and adaptation is equi-
valent to completely retraining the system. This paper briefly outlines the user
modelling problem and work done at BTexact on an Intelligent Personal Assistant
(IPA) which incorporates a user profile. We go on to describe FILUM, a more
flexible method of user modelling, and show its application to the Telephone As-
sistant and Email Assistant components of the IPA, with tests to illustrate its use-
fulness. An experimental testbed based on the iterated prisoner’s dilemma, which
allows the generation of unlimited data for learning or testing, is also proposed.

Introduction

User modeling is a key technology in increasing the effective use of computers
and information appliances, and the next generation of consumer goods requires
more sophistication in user-modeling and intelligent help systems to cope with a
less technologically literate user base. For example, an integrated home informa-
tion / entertainment system (computer, VCR, TV, hi-fi, etc) should be able to sug-
gest TV/video choices based on past preferences, and automatically record pro-
grammes judged to be interesting to a user. With the increased access to informa-
tion arising from the web and integration of digital TV and computer networking,
this area of intelligent consumer goods is an extremely important next step.

2 Soft Computing and User Modeling

We define user modeling to be the provision of a software sub-system able to
observe and predict the actions of a user (from a limited set of possibilities), with
the aim of improving the overall interaction between user and system. This is a
relatively “soft” definition, as the quality of interaction is almost always a subjec-
tive judgment, and it is therefore difficult to discuss the success (or otherwise) of
user modeling.
We can recognise a strongly growing strand of interest in user modelling
arising from research into intelligent interfaces. In this context, we can identify
three different outcomes of user modelling:
t Changing the way in which some fixed content is delivered to the user.
C Changing the content that is delivered to the user.
C Changing the way in which the device is used.
Each of these is discussed in turn below.
The first relates more to the device that is displaying content to a user. For ex-
ample, a WAP browser must restrict graphical content. There is little room for
user likes and dislikes, although [12] describe a system which implements differ-
ent interfaces for different users on desktop systems. Those who have more diffi-
culty navigating through the system use a menu-based interface whereas those
with a greater awareness of the system contents are given an interface using a
number of shortcut keys.
The second category—improving (or changing) information content—is per-
haps the most common. Examples abound in internet-related areas, with applica-
tions to
t Deliver only “interesting” news stories to an individual’s desktop. The point-
cast news delivery systems are a first step (e.g.
www.pointcast.com/products/pcn/ and cnn.com/ads/advertiser/pointcast2.0/);
see also [11] and IDIoMS [13].
s Identify interesting web pages—for example Syskill &Webert [25] uses an in-
formation-theoretic approach to detect “informative” words on web pages.
These are used as features, and user ratings of web pages (very interesting, in-
teresting, not interesting, etc.) creates a training data set for a naive Bayesian
classifier. A similar approach can be used for the retrieval of documents from
digital libraries, using term frequency/inverse document frequency [31] to se-
lect keywords and phrases as features. A user model can be constructed in
terms of these features, and used to judge whether new documents are likely to
be of interest.
b Remove unwanted emails – see [18]or [32, 33] for example. (e.g. [5-7] as well
as earlier work listed in [2])
The problem of “information overload” from email was identified as far back as
1982
“in current message systems, the message remains uninterpreted … The system delivers
the message but does not manage the messages …. In order to enhance their functionality,
message systems have to interpret, at least partially, the messages they handle …”
[8] quoted in [1]. The latter authors also noted that

Introduction 3

“information inundation may cause information entropy, when “incoming messages are
not sufficiently organized by topic or content to be easily recognized as important”
With the incorporation of powerful embedded computing devices in consumer
products, there is a blurring of boundaries between computers and other equip-
ment, resulting in a convergence to information appliances or information devices.
Personalisation, which is equivalent to user modelling, is a key selling point of
this technology—for example, to personalise TV viewing (www.tivo.com, 1999):
“With TiVo, getting your favorite programs is easy. You just teach it what shows you
like, and TiVo records them for you automatically. As you’re watching TV, press the
Thumbs Up or Thumbs Down button on the TiVo remote to teach TiVo what you like As
TiVo searches for shows you’ve told it to record, it will also look for shows that match your
preferences and get those for you as well...”
Sony have implemented a prototype user modelling system [37] which predicts a
viewing timetable for a user, on the basis of previous viewing and programme
classification. Testing against a database of 606 individuals, 108 programme cat-
egories and 45 TV channels gave an average prediction accuracy of 60-70%. We
will not discuss social or collaborative filtering systems here. These are used to re-
commend books (e.g. amazon.com), films, and so on, and are based on clustering
the likes and dislikes of a group of users.
The third category - changing the way in which the device is used - can also be
illustrated by examples. Microsoft’s Office Assistant is perhaps the best known
example of user modelling, and aims to provide appropriate help when required,
as well as a “tip of the day” that is intended to identify and remedy gaps in the
user’s knowledge of the software. The Office Assistant was developed from the
Lumiere [16] project, which aimed to construct Bayesian models for reasoning
about the time-varying goals of computer users from their observed actions and
queries. Although it can be argued that the Office Assistant also fits into the previ-
ous category (changing the content delivered to the user), its ultimate aim is to
change the way the user works so that the software is employed more effectively.
The system described by [20] has similar goals but a different approach. User
modelling is employed to disseminate expertise in use of software packages (such
as Microsoft Word) within an organisation. By creating an individual user model
and comparing it to expert models, the system is able to identify gaps in know-
ledge and offer individualised tips as well as feedback on how closely the user
matches expert use of the package. The key difference from the Office Assistant is
that this system monitors all users and identifies improved ways of accomplishing
small tasks; this expertise can then be spread to other users. The Office Assistant,
on the other hand, has a static view of best practice.
Hermens and Schlimmer [14] implemented a system which aided a user filling
in an electronic form, by suggesting likely values for fields in the form, based on
the values in earlier fields.
The change in system behaviour may not be obvious to the user. Lau and Hor-
vitz [19] outline a system which uses a log of search requests from Yahoo, and
classifies users’ behaviour so that their next action can be predicted using a
Bayesian net. If it is likely that a user will follow a particular link, rather than re-


fining or reformulating their query, then the link can be pre-fetched to improve the
perceived performance of the system. This approach generates canonical user
models, describing the behaviour of a typical group of users rather than individual
user models.
There are two key features in all these examples:
T the aim is to improve the interaction between human and machine. This is a
property of the whole system, not just of the machine, and is frequently a sub-
jective judgement that can not be measured objectively.
j the user model must adapt in the light of interaction with the user.
t Additionally, it is desirable that the user model
A be gathered unobtrusively, by observation or with minimal effort from the user.
b be understandable and changeable by the user - both in terms of the knowledge
held about the user and in the inferences made from that knowledge.
h be correct in actions taken as well as in deciding when to act.

User models—Learning, Adaptivity and Uncertainty

The requirement for adaptation puts user modelling into the domain of machine
learning (see [17] and [36]). A user model is generally represented as a set of at-
tribute-value pairs—indeed the W3C proposals [34] on profile exchange recom-
mend this representation. This is ideal for machine learning, as the knowledge rep-
resentation fits conveniently into a propositional learning framework. To apply
machine learning, we need to gather data and identify appropriate features plus the
desired attribute for prediction. To make this concrete, consider a system which
predicts the action to be taken on receiving emails, using the sender’s identity and
words in the title field. Most mail readers allow the user to define a kill file, spe-
cifying that certain emails may be deleted without the user seeing them. A set of
examples might lead to rules such as
if title includes $ or money then action = delete
if sender = boss then action = read, and subsequently file
if sender = mailing list then action = read and subsequent-
ly delete
This is a conventional propositional learning task, and a number of algorithms ex-
ist to create rules or decision trees on the basis of data such as this [4, 5, 7, 27, 28].
Typically, the problem must be expressed in an attribute-value format, as above;
some feature engineering may be necessary to enable efficient rules to be induced.
Rule-based knowledge representation is better than (say) neural nets due to better
understandability of the rules produced - the system should propose rules which
the user can inspect and alter if necessary. See [24] for empirical evidence of the
importance of allowing the user to remain in control.
One problem with propositional learning approaches is that it is difficult to ex-
tract relational knowledge. For example:

Introduction 5

if several identical emails arrive consecutively from a
list server, then delete all but one of them

Also, it can be difficult to express relevant background knowledge such as:
if a person has an email address at acme.com then that per-
son is a work colleague

These problems can be avoided by moving to relational learning, such as inductive
logic programming [23], although this is not without drawbacks as the learning
process becomes a considerable search task.
Possibly more serious issues relate to the need to update the user model, and to
incorporate uncertainty. Most machine learning methods are based on a relatively
large, static set of training examples, followed by a testing phase on previously
unseen data. New training examples can normally be addressed only by restarting
the learning process with a new, expanded, training set. As the learning process is
typically quite slow, this is clearly undesirable. Additionally in user modelling it is
relatively expensive to gather training data - explicit feedback is required from the
user, causing inconvenience. The available data is therefore more limited than is
typical for machine learning.
A second problem relates to uncertainty. User modeling is inherently uncertain
—as [15] observes, “Uncertainty is ubiquitous in attempts to recognize an agent’s
goals from observations of behavior,” and even strongly logic-based methods such
as [26] acknowledge the need for “graduated assumptions.” There may be uncer-
tainty over the feature definitions. For example:
if the sender is a close colleague then action = read very
soon

where close colleague and very soon are fuzzily defined terms) or over the applic-
ability of rules. For example:
if the user has selected several options from a menu and
undone each action, then it is very likely that the user
requires help on that menu

where the conclusion is not always guaranteed to follow.
It is an easy matter to say that uncertainty can be dealt with by means of a fuzzy
approach, but less easy to implement the system in a way that satisfies the need for
understandability. The major problem with many uses of fuzziness is that they rely
on intuitive semantics, which a sceptic might translate as “no semantics at all.” It
is clear from the fuzzy control literature that the major development effort goes
into adjusting membership functions to tune the controller. Bezdek [9, 10] sug-
gests that membership functions should be “adjusted for maximum utility in a giv-
en situation.” However, this leaves membership functions with no objective mean-
ing—they are simply parameters to make the software function correctly. For a
fuzzy knowledge based system to be meaningful to a human, the membership
functions should have an interpretation which is independent of the machine oper-
ation—that is, one which does not require the software to be executed in order to
determine its meaning. Probabilistic representations of uncertain data have a


strictly defined interpretation, and the approach adopted here uses Baldwin’s mass
assignment theory and voting model semantics for fuzzy sets [3, 8].

The Intelligent Personal Assistant

BTexact’s Intelligent Personal Assistant (IPA) [1, 2] is an adaptive software sys-
tem that automatically performs helpful tasks for its user, helping the user achieve
higher levels of productivity. The system consists of a number of assistants spe-
cialising in time, information, and communication management:
c The Diary Assistant helps users schedule their personal activities according to
their preferences.
t Web and Electronic Yellow Pages Assistants meet the user’s needs for timely
and relevant access to information and people.
a The RADAR assistant reminds the user of information pertaining to the current
task.
t The Contact Finder Assistant puts the user in touch with people who have sim-
ilar interests.
i The Telephone and Email Assistants give the user greater control over incom-
ing messages by learning priorities and filtering unwanted communication.
As with any personal assistant, the key to the IPA’s success is an up-to-date un-
derstanding of the user’s interests, priorities, and behaviour. It builds this profile
by tracking the electronic information that a user reads and creates over time—for
example, web pages, electronic diaries, e-mails, and word processor documents.
Analysis of these information sources and their timeliness helps IPA understand
the users personal interests. By tracking diaries, keyboard activity, gaze, and
phone usage, the IPA can build up a picture of the habits and preferences of the
user.
We are particularly interested in the Telephone and E-mail assistants for com-
munication management, used respectively for filtering incoming calls and priorit-
ising incoming e-mail messages. The Telephone Assistant maintains a set of prior-
ities of the user’s acquaintances, and uses these in conjunction with the caller’s
phone number to determine the importance of an incoming call. The Email Assist-
ant computes the urgency of each incoming message based on its sender, recipi-
ents, size and content. Both assistants use Bayesian networks for learning the in-
tended actions of the user, and importantly, the system continually adapts its beha-
viour as the user’s priorities change over time.
The telephone assistant handles incoming telephone calls on behalf of the user
with the aim of minimising disruption caused by frequent calls. For each incoming
call, the telephone assistant determines whether to interrupt the user (before the
phone rings) based on the importance of the caller and on various contextual
factors such as the frequency of recent calls from that caller and the presence of a
related entry in the diary (e.g. a meeting with the caller). When deciding to inter-
rupt the user, the telephone assistant displays a panel indicating that a call has ar-

Introduction 7

rived; the user has the option of accepting or declining to answer the call. The tele-
phone assistant uses this feedback to learn an overall user model for how the user
weights the different factors in deciding whether or not to answer a call. Although
this model has been effective, its meaning is not obvious to a user, and hence it is
not adjustable. To address this issue, the FILUM [21, 22] approach has been ap-
plied to the telephone and email assistants.

Assumptions for FILUM

We consider any interaction between a user and a software or hardware system in
which the user has a limited set of choices regarding his/her next action. For ex-
ample, given a set of possible TV programmes, the user will be able to select one
to watch. Given an email, the user can gauge its importance and decide to read it
immediately, within the same day, within a week, or maybe classify it as unim-
portant and discardable. The aim of user modelling is to be able to predict accur-
ately the user’s decision and hence improve the user’s interaction with the system
by suggesting or making such decisions automatically. Human behaviour is not
generally amenable to crisp, logical modelling. Our assumption is that the limited
aspect of human behaviour to be predicted is based mainly on observable aspects
of the user’s context—for example, in classifying an email the context could in-
clude features such as the sender, other recipients of the message, previously re-
ceived messages, current workload, time of day, and so on. Of course, there are
numerous unobservable variables - humans have complex internal states, emo-
tions, external drives, and so on. This complicates the prediction problem and mo-
tivates the use of uncertainty modelling—we can only expect to make correct pre-
dictions “most” of the time.
We define a set of possible output values
B = {b1, b2, …, bj},
which we refer to as the behaviour, and a set of observable inputs
I = {i1, i2, …, im}.
Our assumption is that the n+1th observation of the user’s behaviour is predict-
able by some function of the current observables and all previous inputs and beha-
viours.
bn+1 = f(I1, b1, I2, b2, … In, bn, In+1)
The user model, including any associated processing, is equivalent to the function
f. This is assumed to be relatively static; within FILUM, addition of new proto-
types would correspond to a major change in the function.
We define a set of classes implemented as Fril++ [6, 29] or java programs.
C = {c1, c2, … ck},


A user model is treated as an instance that has a probability of belonging to each
class according to how well the class behaviour matches the observed behaviour
of the user. The probabilities are expressed as support pairs, and updated each
time a new observation of the user’s behaviour is made.
We aim to create a user model m, which correctly predicts the behaviour of a
user. Each class ci must implement the method Behaviour; giving an output in B
(this may be expressed as supports over B). Let Sn (m ∈c i ) be the support for the
user model m belonging to the ith class before the nth observation of behaviour.
Initially,
S1(m ∈ ci ) = [0, 1]
for all classes ci, representing complete ignorance.
Each time an observation is made, every class makes a prediction, and the sup-
port for the user model being a member of that class is updated according to the
predictive success of the class :

n(c (h r= )
S ) . v+ b
× ∈ c an=
+
mSeo
Bu n
i 1 +
( ) n i i
S ∈
+ c
nm
1 i= 1
(0)
n+1
where S (ci .Behaviourn +1 == bn +1) represents the (normalised) support for class ci
predicting the correct behaviour (from the set B) on iteration n+1.
The overall model behavior is predicted by multiplying the support for each
prototype by the support for each behavior prediction made by that prototype, and
then taking the best support over all prototypes.
i.e.
c v a ∈ er)
P erm c B=
re i
e h
da
iB
t du x o = () h = ( Sv )
S ×a b
m (i
co
. u
∈
C∈
c B
,
b

where B is the set of possible behaviors and P is the set of user prototypes.
For example, take the problem of predicting the outcome of a biased coin
which always lands on heads (this corresponds to a “user”, with behaviors “heads”
and “tails”) and three prototypes P1, P2, P3 which are simple probabilistic pro-
grams predicting heads with probabilities 0.4, 0.5 and 0.6 respectively (we will
work with probabilities rather than supports to simplify the example). After sever-
al iterations (tosses), their success rates in correctly predicting the outcome (i.e.
support for m ∈ Ci) will be as shown in the table below,

prototype S(m ∈ C i) prediction weighted prediction
C1 0.4 heads 0.4 heads 0.16
tails 0.6 tails 0.24

Introduction 9

The next prediction would be heads, as this has the highest weighted support
(0.36 from prototype C3). This is the prediction we would intuitively expect, and
would also be the prediction of the weighted sum of supports (Heads :
0.16+0.25+0.36 = 0.77, Tails : 0.24+0.25+.0.24 = 0.73).
Note that if we add a fourth prototype which predicts heads 10% of the time,
the user model behaves as before :

prototype S(m ∈ C i) prediction weighted prediction

The prototypes in the user model are more inclined towards “tails”, so that tak-
ing a weighted combination of predictions would not be an accurate user model. In
practice it may be impossible to determine whether a set of prototypes is biased to-
wards any particular behavior – the whole point of the FILUM approach is that the
prototypes provide knowledge-based estimations of the effects of external factors,
because adequate statistics are not available.

Updating Support

The user model is treated as a partial instance of all prototype classes, with its
degree of membership in each class determined by the accuracy with which the
prototype predicts the correct behavior, as shown in Eq. 1. Note that this does not
discriminate against a prototype which gives high support to all possible behaviors
– to take an extreme case, allocating a support of (1 1) to every possibility. Thus in
addition to requiring each prototype to give the support for each behavior, they
must also predict a single behavior. If this is correct, support is updated as above;
if it is wrong, the updating support is (0 0). Prototypes are allowed to return
FILUMPrototype.NOPREDICTION in cases where they are not applicable.
To illustrate this process, consider three prototypes – one which always predicts
correctly, one which always predicts wrongly and a third which is alternately cor-
rect and incorrect. The supports for these prototypes will evolve as shown in Fig
1.
Clearly as n becomes large, supports change relatively slowly. [30] discuss an
alternative updating algorithm which is weighted in favour of more recent beha-
viour, particularly unexpected actions. The accuracy of the user model at any stage
is the proportion of correct predictions made up to that point—this metric can eas-
ily be changed to use a different utility function, for example, if some errors are
more serious than others.


1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6 7 8 9 10 11 12

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6 7 8 9 10 11 12

1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6 7 8 9 10 11 12

Fig.1 : evolution of support over 12 iterations for three user prototypes. The dark bars
show the necessary support, diagonal stripes represent uncertainty and light horizontal
shading represents support against the prototype predicting correctly. The top graph
shows a prototype which always predicts correctly and the support evolves towards 1;
the second graph shows a prototype which always predicts incorrectly, and the support
evolves towards 0. The third shows a prototype which alternates between correct and in-
correct predictions, and the support evolves towards 0.5

Introduction 11

Testing

In order to test user modeling approaches, data is needed. This can either be gath-
ered from a real situation or generated using an artificial model problem. The n-
player iterated prisoner’s dilemma described below provides a method of generat-
ing data for user modeling experiments. It should be emphasized that the game it-
self is of no interest here; it is used purely as a platform for generating data to test
user modeling approaches.
The Prisoner’s Dilemma is a well-known example of a non-zero sum game in
game theory. As an illustration, consider two prisoners who are jointly charged
with a crime for which the standard sentence is five years. They are separately of-
fered a deal whereby they can stay silent (co-operate with their partner) or defect,
providing evidence to implicate their partner. Each can choose to co-operate or de-
fect, and the reduction in their sentences according to their joint behavior is given
in the following table:

player1 / player2 co-operate defect
co-operate 3/3 0/5
defect 5/0 1/1

From each individual’s point of view, the rational strategy is to defect; collec-
tively the best solution is for both to co-operate.
The iterated version extends the game to a sequence of interactions, where each
player has access to the history of interactions. The n-player version considers
more than two individuals. In each round, each pair of players participates in a
pairwise interaction as above.
In all cases, each player aims to maximize their own score. There is an incen-
tive to co-operate (e.g. the payoff from three co-operative interactions, c-c, c-c, c-c
will be 9 whereas the payoff from one “exploitative” interaction and two mutual
defections (d-c, d-d, d-d) will only be 7).
The iterated prisoner’s dilemma or IPD [1] has been used to explain arms race
escalation, the formation of economic cartels, and evolutionary biology, as well as
acting as a test bed for multi-agent systems [10] and evolutionary programming. It
can be shown that there is no optimum strategy, as much is dependent on the envi-
ronment. For example if all other players co-operate irrespective of one’s own ac-
tions, the optimum strategy is to defect. If all players respond by echoing one’s
last move (the “tit-for-tat” strategy) then the optimum strategy is to co-operate.
The game was widely publicized in computer tournaments e.g. [13,20] in which
co-operative strategies tended to do best. In particular successful strategies are
c “nice” in that they do not defect without provocation
“ “responsive” in that they punish provocation (i.e. defection by the other player)
“ “forgiving” in that they will attempt to co-operate again after punishing an op-
ponent


n-player IPD as a Test Bed for User Modeling

The n-IPD is a good test bed for user modeling as it is possible to generate as
much data as necessary and the true behavior can be obtained in each case, so that
it is possible to get an objective evaluation of the predictive accuracy of the user
model.
The aim is to replace a selected player in an n-player tournament by a user
model, which behaves in the same way as the player without knowledge of or ac-
cess to the internal structure of the player, i.e. without knowing the code govern-
ing the player’s behavior. The user model has access to the interaction history of
the selected player. There is a close analogy to the situation in user modeling ap-
plications, where the previous behavior of a user is known without any detailed
understanding of the “algorithm” that led to the decision.
There is no intention to judge whether a given strategy is successful or to opti-
mize a strategy. At each iteration, the user model of a player P simply makes a
prediction of P’s behavior in an interaction with each other player in the tourna-
ment.

Experiments

A Fril++[3,5,6] -based system was developed to run n-IPD tournaments, allowing
a choice of strategies to be included in the environment. The number of players
using each strategy could also be specified. Examples of strategies are:
u trust - always co-operates
defect - always defects
tft (tit-for-tat) initially co-operates, subsequently echoes whatever opponent did
last time
l rand - random (50-50) choice
crand - random (75-25) choice biased towards co-operation
drand - random (25-75) choice biased against co-operation
tftt - (tit-for-two-tats) co-operates unless two consecutive defections from oppo-
nent
n stft - sneaky tit-for-tat initially defects then echoes last opponent response
massret (massive retaliation) - co-operates until a defection, then defects
against that opponent forever
a other responsive strategies - a method must be supplied to determine initial be-
haviour, subsequently co-operate unless the number of defections by opponent
in a given number of previous interactions exceeds some threshold, in which
case defect (tit for tat, tit for two tats, massret are all subtypes of this strategy).
Examples included in the tournaments are
E respc - co-operate unless all of the opponent’s last 6 responses were d
d respd - defect unless all of the opponent’s last 6 responses were c
Figure 2 shows the results from two sample tournaments, plotting the average re-
ward per interaction for selected player in the population. If all interactions are co-

Introduction 13

operative, then the average will be 3; similarly if a player manages to exploit ev-
ery other player then its average reward would be 5. In Figure 1 (a) there are too
few responsive players, and too many that co-operate unconditionally. The “defect
every time” strategy is able to exploit the other players and maintain a clearly bet-
ter overall average.


In Figure 2 (b) the tournament contained a similar number of co-operative players
but more were responsive, and withdrew co-operation from the the “defect every
time” strategy. The latter does worse than the others, despite its initial success.
The precise details are not important, other than to note that the average score is
higher in more co-operative environments, and that the best strategy in one envi-
ronment is not necessarily the same as the best strategy in another. Such observa-
Exploitable environment

5

4.5

4
tft1
3.5
defect1
3 drnd1
stft1
2.5
Respd
2

1.5

1

Iteration

Responsive and co-operative environment

5

4.5

4

3.5 tft1
trust1
3
defect1
2.5 Respc
2

1.5

1

Iteration

Fig. 2: Average reward per interaction for various players in two n-IPD tournaments

(a) exploitable environment, where the top line represents a player defecting on every move, and

(b) responsive environment where the defecting strategy scores highly at first but then performs
poorly. Twenty strategies were represented in each tournament, drawn from those listed in Sec-
tion 7. It is possible for the same strategy to be included several times as different “players” in
the tournament.

Introduction 15

tions are in line with other n-IPD tournaments and help to verify that the software
performs correctly.
From the point of view of user modeling, we aim to reproduce the behavior of a
player by means of some simple prototypes. The properties of successful players
observed in tournaments are niceness, responsiveness, and forgiveness. In order to
roughly model these characteristic behaviors, the four prototypes exhibit the fol-
lowing patterns:
l trusting - always co-operates
t defective - always defects
d responsive - identified by echoing the opponents last move
r provocative - identified by defecting when the opponent co-operated last time
(this can be viewed as the negation of nice)
Note that there is no prototype to explicitly identify “forgiving” behavior, al-
though the “responsive” prototype will effectively detect it.
Thus if we regard the interaction between two specified players P and Q as a
sequence of pairs of actions (pi, qi), we are looking for instances of :

(c, _) to support P belonging to the co-operative prototype
(d, _) to support P belonging to the unco-operative prototype
(c, d), (d, _) or to support P belonging to the responsive prototype
(d, c), (c, _)
(c, c), (d, _) to support P belonging to the provocative prototype

With a history of interactions generated by an n-IPD tournament, we can use
these prototypes to model different players. Table 1 shows selected user models
after 12 iterations of a 20 player tournament. The user models have converged af-
ter 12 iterations, i.e. the supports for the models belonging to each class do not
change significantly after this iteration.

Table 1 : Protoype supports in user models derived from 12 iterations of an n-IPD tourna-
ment. Column headings refer to prototypes Trusting, Responsive, Provocative and Defec-
tive as defined above
Player T R P D
co-op [0.9 1] [0, 1] [0. 0.09] [0, 0.09]
uncoop [0, 0.09] [0, 1] [0. 0.09] [0.9 1]
tit-for-tat [0.6,0.7] [0.3,1] [0,0.5] [0.3,0.4]
random [0.4,0.5] [0.3,0.6] [0.2,0.7] [0.5,0.6]
respd [0.2,0.3] [0.1,0.8] [0.1,0.9] [0.7,0.8]
Prediction success rates vary between 40-60% for the random strategies, and
80-95% for the others.
The support for class membership is determined purely by the success of the
class rule in predicting the behavior of the user. Thus a player showing a strategy
of “always defect” appears as highly unco-operative but has [0, 1] support for


membership in the “provocative” class, as the player never exhibits the provoca-
tive behavior pattern of mutual co-operation followed by a defection.
The models for the selected player are derived from all of the player’s interac-
tions; within each model, there are models for the other players derived from their
interactions with the selected player.
Overall predictive success rates are good although the random strategies are
difficult to predict, as would be expected. As a rule of thumb, if a user model is
giving a success rate of less than 60% then a new prototype is required, either by a
human expert or using an inductive logic-type approach to generate prototypes
which predict the observed behavior more accurately. Note that these new proto-
types need only explain the behavior in a subset of cases; they can give uncertain
support for all outcomes when the prototype is not applicable.
The success rate of a user model can easily be calculated by comparing the pre-
dicted and observed behavior of the user. Clearly the user model changes with
each new observation, and there is very little overhead in updating the user model.
This approach depends on having a “good” set of prototypes, which are able to
give a reasonable coverage of possible user behavior. It is assumed that a human
expert is able to provide such a set; however, it is possible that new prototype be-
haviors could be generated by techniques such as inductive logic programming.
This is an interesting avenue for future research.

4.2 User Models in the Telephone Assistant

The FILUM approach has also been applied to the prediction of user behaviour in
the telephone assistant. The following assumptions have been made:
T The user model must decide whether to divert the call to voicemail or pass it
through to be answered.
t The user is available to answer calls.
T Adaptive behaviour is based on knowing the correct decision after the call has
finished.
f A log of past telephone activity and the current diary are available
A The identity of all callers is known.
A sample set of user prototypes is shown in Table 2.

Introduction 17

Table 2. User Prototypes for the telephone assistant

Prototype Identifying Characteristic Behaviour
Talkative none always answer
Antisocial none always divert to voicemail
Interactive recent calls or meetings involving this answer
caller
Busy small proportion of free time in next answer if caller is brief, other-
working day (as shown by diary) wise divert to voicemail
Overloaded small proportion of free time in next divert to voicemail
working day (as shown by diary)
Selective none answer if caller is a member of
a selected group, else divert to
voicemail
Regular large proportion of calls answered at answer if this is a regular time
particular times of the day e.g. early
morning

This approach assumes that all activities are planned and recorded accurately in an
electronically accessible format. Other ways of judging a user’s activity would be
equally valid and may fit in better with a user’s existing work pattern - for ex-
ample the IPA system investigated the use of keyboard activity, gaze tracking and
monitoring currently active applications on a computer. There is a need to model
callers using a set of caller prototypes, since a user can react in different ways to
different callers in a given set of circumstances. For example, the phone rings
when you are due to have a meeting with the boss in five minutes. Do you answer
if (a) the caller is the boss or (b) the caller is someone from the other side of the
office who is ringing to talk about last night’s football results while waiting for a
report to print. The sample set of caller prototypes is shown in Table 3.
The user and caller prototypes are intended to illustrate the capabilities of the
system rather than being a complete set; it is hoped that they are sufficiently close
to real behaviour to make detailed explanation unnecessary.
Terms in italic are either fuzzy definitions that can be changed to suit a user.
Note that support pairs indicate the degree to which a user or caller satisfies a par-
ticular prototype - this can range from uncertain (0 1) to complete satisfaction (1
1) or its opposite (0 0), through to any other probability interval.

Table 3. Caller Prototypes

Prototype Identifying Characteristic
Brief always makes short calls to user
Verbose always makes long calls to user
Frequent calls user frequently
Reactive calls following a recent voicemail left by user
Proactive calls prior to a meeting with user


A sample diary is shown in Figure 2. Note that the diary is is relatively empty
at the beginning and end of the week but relatively full in the middle of the week.
The busy and overloaded prototypes are written to be applicable when there is a
small proportion of free time in the immediate future, that is, during the latter part
of Tuesday and Wednesday.

5
4 1 design_review
activity

3 2 seminar
2 3 research
1
4 programming
0 52 104 156 208 260
5 home
time (15 minute intervals)

Fig. 2. Sample of diary. The window for the working day has been defined as 7:00 am -
8:00 pm, and diaried activities for each fifteen minute period within the window are shown;
unassigned slots represent free time which can be used as appropriate at the time.

1

0.9

0.8

0.7

0.6

Individual
0.5
Cumulative

0.4

0.3

0.2

0.1

0
1/10/01 1/10/01 2/10/01 2/10/01 3/10/01 3/10/01 4/10/01 4/10/01 5/10/01 5/10/01 6/10/01
0:00 12:00 0:00 12:00 0:00 12:00 0:00 12:00 0:00 12:00 0:00

Fig. 3. Performance of the user model on individual calls (triangles, correct prediction =1,
incorrect prediction = 0) and as a cumulative success rate (diamonds, continuous line).
Fig 3 shows the success rate of the user model in predicting whether a call should
be answered or diverted to voicemail. The drop in performance on the second day
occurs because the busy and overloaded prototypes become active at this time, due
to the full diary on Wednesday and Thursday. It takes a few iterations for the sys-
tem to increase the membership of the user model in the busy and overloaded
classes; once this has happened, the prediction rate increases again.
The necessary and possible supports for membership of the user model in the
busy class is shown in Fig 4, where the evolution of support can be seen on the
third and fourth days where this prototype is applicable. At the start, the identify-
ing characteristics (full diary) are not satisfied and support remains at unknown (0

Introduction 19

1

0.9

0.8

Busy
0.7
Busy
0.6

0.5

0.4

0.3

0.2

0.1

0
1/10/01 1/10/01 2/10/01 2/10/01 3/10/01 3/10/01 4/10/01 4/10/01 5/10/01 5/10/01 6/10/01
0:00 12:00 0:00 12:00 0:00 12:00 0:00 12:00 0:00 12:00 0:00

Fig. 4. Evolution of support pair for the “busy” prototype in the user model. The lower and
upper lines show respectively the necessary and possible support for the observed behavior
matching this prototype. At the start, the identifying characteristics (full diary) are not
satisfied and support remains at unknown (0 1). In the middle of the week, conditions are
satisfied. Initially the user does not behave as predicted by this prototype and possible support
drops (i.e. support against increases); subsequently, the user behaves as predicted and
necessary support increases. At the end of the week, once again the identifying characteristics
are not satisfied and the prototype makes no predictions.

1). In the middle of the week, conditions are satisfied. Initially the user does not
behave as predicted by this prototype and possible support drops (i.e. support
against increases); subsequently, the user behaves as predicted and necessary sup-
port increases. At the end of the week, once again the identifying characteristics
are not satisfied.
By repeating the week’s data, there is relatively little change in the support
pairs—this is suggestive that the learning has converged, although additional test-
ing is necessary. Evolution of caller models can also be followed within the sys-
tem, and good convergence to a stable caller model is observed. It should be em-
phasised that the supports for each prototype can be adjusted by the user at any
stage. The user modelling software has been tested in several diary and call log
scenarios, with good rates of prediction accuracy. Further testing is needed to in-
vestigate the user model evolution over longer periods.

User Models in the Email Assistant

The FILUM approach has also been also applied to the prediction of user beha-
viour in the email assistant. In many respects, this is a simpler problem than the
telephone assistant – for example, in the latter there was a need to create caller


models to predict the likely duration of a call. In the email assistant, this is no
longer true as the required data is available from the fields of the email message,
and only the user needs to be modeled. The aim is to predict the user’s classifica-
tion of emails into the following categories:
B = {READIMMEDIATE, READ24HOURS, READ5DAYS, READ4WEEKS, READ-
NEVER}
on the basis of the message content and header fields i.e. to, from, date, time, subject and
length.
The set of behaviors could easily be extended to more sophisticated actions
such as “read this message now, then file in folder X” or “read now and answer
within 24 hours”. Rather than automatically moving emails to the predicted folder,
a list of most likely destinations could be provided, with a facility for the user to
enable automatic execution once s/he was happy with the accuracy of the system.
Alternatively we could redirect mail to a different device, when appropriate. If a
user is only available by telephone, a short and high priority (read immediately)
message could be converted to speech and relayed to the user; if a handheld device
had a greater capacity then all high priority messages could be redirected.
The set of behaviours was chosen to match those used in previous versions of
the email assistant. The sample set of user prototypes is shown in Table 4.

Introduction 21

Table 4. User Prototypes for the email assistant

Prototype Identifying Characteristic Behaviour

Always readNever none always classify as readNever
Always readImmed none always classify as readImmediate
Always 24Hr none always classify as read24Hours
Always 5Days none always classify as read5Days
Always 4Weeks none always classify as read4Weeks
Always uncertain none equal support (0 1) for all beha-
viours
PreviousBehav previous emails received from repeat behaviour on last email from
this sender this sender
FixedTimes current time matches set inter- classify as readImmediate if current
vals time is in the fuzzy intervals 8:00 –
10:00 or 16:00 – 18:00; otherwise
inherit classifiaction method from
PreviousBehav
Len/keywords none Uses a fuzzy “acceptable” message
length and fixed fuzzy sets of
words to identify mails to be read
immediately (“good” keywords) or
never (“bad” keywords). Otherwise
returns (0 1) for all behaviours.

All testing was carried out on a data set of 94 email messages. Several different
user behaviours were simulated on this set of emails, using different “verdict” files
as shown in Table 5.

Table 5. Success rates of user model in email prediction against different scenarios


Short description categorisation method % Accuracy of user model
after all emails processed

verdictAll24Hours all messages categorised as “Read 99
in 24 hours”

verdictAll4W all messages categorised as “Read 99
in 4 weeks”

verdictAll5Day all messages categorised as “Read 99
in 5 days”

verdictAllImmed all messages categorised as “Read 99
immediately”

verdictAllNo all messages categorised as “Read 99
never”

verdictByHand manually assigned categories 80

verdictFromName Sender contains 83
“bt.co” -> immediate
sales or market -> never
yahoo or hotmail -> 5 days
otherwise 24 hours

verdictFromTime if hour = 81
8,9,16,17then immediate
7,10,18,19 then 24 hours
otherwise random

verdictRandom random categorisation 16

verdictRandom2 random categorisation 26

NB accuracy is given as the proportion of correct predictions. In some cases
this involves a random choice, so results may vary slightly from run to run. Evolu-
tion of prediction accuracy as emails are presented is also shown below for the
sixth and seventh cases in the table. The first five and last two cases are to test the
user model’s behavior in extremes; the remaining three cases are more realistic.

Introduction 23

Prediction success

100

90

80

70

60 Predicted
50 Actual

40 Success Rate

30

20

10

0
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93

User model - support for prototypes

1

0.8

0.6
Uncertainty
Nec support
0.4

0.2

0
es
av
s

Hr
ks

ds
r

in
ed

ay

im
ve

eh
24
ee
rta

or
5D
m

dT
Ne

sB

yw
4W
ce
Im

ys

xe
ad

ys

iou
un

ke
wa
ad

Fi
ys
re

wa

ev

n/
re

Al
ys

wa
ys

Le
Al

Pr
wa
ys
wa

Al
wa

Al
Al

Al

Fig 5 Prediction success and user model for verdicts according to sender name

It is interesting to note that whilst success rates are roughly the same in these
cases, the user models are very different, e.g. the prototype based on message
length and keywords (FILUMUserPrototypeLW) has very little support in the ver-
dictFromName case, but high support in the verdictByHand scenarios (and ver-
dictFromTime, although this is not shown.


Because the prototypes almost always return a prediction, there is little uncer-
tainty in the user models - with more sophisticated prototypes, there would be a
greater difference between necessary and possible supports.
The predicted and actual behaviours are shown on the success plots - 0 corre-
sponds to READIMMEDIATE, … 40 corresponds to READNEVER.
Prediction success

100

90

80

70

60 Predicted
50 Actual
40 Success Rate

30

20

10

0
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91

User model - support for prototypes

1

0.8

0.6
Uncertainty
Probability
0.4

0.2

0
es
ks

v
s
ed

r
er

ds
in

ha
H
ay

im
ee
rt a
ev

24
m

or
Be
5D

dT
4W
Im

yw
N

ce

s

us
ad

xe
ay
ad

s
un

ke
io
ay
s
re

Fi
w
re

ay

n/
ev
s

w

Al
s

ay

Le
w
s

Al

Pr
ay

ay

Al
w
w

Al
w
Al

Al

Fig 6 Prediction success and user model for realistic verdicts (generated by hand)

Introduction 25

Summary

Many research projects have considered the use of machine learning for user mod-
elling. In general, the problems identified include
e collection of training data
c length of time to train
l changes in profile cannot be incorporated without costly retraining
The work reported here extends previous work on FILUM and addresses the sec-
ond and third issues listed above.
The aim of user modelling is to increase the quality of interaction—this is al-
most always a subjective judgement, and it can be difficult to discuss the success
(or otherwise) of user modelling. We have developed an experimental testbed
based on the iterated prisoner’s dilemma, allowing generation of unlimited data.
Prediction success rates vary between 80-95% for non-random behaviours in the
testbed, and accuracy of over 80% has been obtained in a series of simulated tests
of the telephone and email assistants.
The user model changes with each observation, and there is very little overhead
in updating the user model. This approach depends on having a “good” set of pro-
totypes, giving reasonable coverage of possible user behaviour. It is assumed that
a human expert is able to provide such a set; however, it is possible that (for ex-
ample) inductive logic programming could generate new prototype behaviours.
This is an interesting avenue for future research.
Finally, there is a need for more rigorous testing – although it is difficult to ob-
tain ‘real” data due to privacy issue.

References

BenMartine.doc

More Related Content

What's hot (16)

Viewers also liked (9)

Similar to BenMartine.doc (20)

More from butest (20)

BenMartine.doc