Content uploaded by Jan Treur

Author content

All content in this area was uploaded by Jan Treur on Feb 23, 2014

Content may be subject to copyright.

AUTHOR COPY

Modelling biased human trust dynamics

1

Mark Hoogendoorn a,*, Syed Waqar Jaffry a,b, Peter-Paul van Maanen a,c and Jan Treur a

a Agent Systems Research Group, VU University Amsterdam, De Boelelaan 1081, 1081 HV Amsterdam,

The Netherlands

E-mail: {mhoogen,swjaffry,treur}@few.vu.nl

b Punjab University College of Information Technology (PUCIT), University of The Punjab,

Shahrah-e-Quaid-i-Azam, Lahore, Pakistan

E-mail: swjaffry@pucit.edu.pk

c Department of Perceptual and Cognitive Systems, Netherlands Organisation for Applied Scientific Research

(TNO), P.O. Box 23, 3769 ZG Soesterberg, The Netherlands

E-mail: peter-paul.vanmaanen@tno.nl

Abstract. Within human trust related behaviour, according to the literature from the domains of Psychology and Social Sci-

ences often non-rational behaviour can be observed. Current trust models that have been developed typically do not incorpo-

rate non-rational elements in the trust formation dynamics. In order to enable agents that interact with humans to have a good

estimation of human trust, and take this into account in their behaviour, trust models that incorporate such human aspects are a

necessity. A specific non-rational element in humans is that they are often biased in their behaviour. In this paper, models for

human trust dynamics are presented incorporating human biases. In order to show that they more accurately describe human

behaviour, they have been evaluated against empirical data, which shows that the models perform significantly better.

Keywords. Trust, biases, modelling, validation

1. Introduction

Within the domain of multi-agent systems, a vari-

ety of trust models have been proposed (e.g., see

[13,14] for an overview). Often, such trust models

are utilized in an environment in which software

agents should make choices based upon their levels

of trust, and hence, such models aim to optimize the

behavior of the agent by using the most appropriate

trust function. An example of such a model is for

instance described in [12]. In situations where soft-

ware agents interact with humans, trust models that

are incorporated in these agents may have a com-

pletely different purpose: to estimate the trust levels

of the human over time, and take that into considera-

tion in its behavior, for example, by providing ad-

vices from other trustees that are trusted more. If this

is the purpose of the trust model, then the model

1 The work presented in this paper is a significant extension by

more than 40% of (Hoogendoorn, Jaffry, Maanen, and Treur,

2011).

* Corresponding author.

should also explicitly incorporate non-rational human

aspects. Examples of models taking into account

various human aspects are [3,7,11].

In the literature in the domain of Psychology and

Social Sciences it has been shown that one important

non-rational aspect within the formation of trust is

the incorporation of biases. Several biases have been

observed whereby the culture bias is one of the most

reported ones. In [20] it is shown that humans from

collectivistic cultures tend to have a bias towards

trusting members that belong to the same group and

avoid the persons from outside the group. In [8] also

a comparison between individualistic and collecti-

vistic cultures is made which shows that the trust of

the members of an individualistic society is less

negatively biased towards persons from outside their

group. Other authors also emphasize the existence of

such a bias in general, e.g., [15]. If the objective of a

computational model of trust is to create a model that

represents human trust in a natural and accurate

manner, such biases need to be taken into account in

the model.

Web Intelligence and Agent Systems: An International Journal 11 (2013) 21–40

DOI 10.3233/WIA-130260

IOS Press

1570-1263/13/$27.50 © 2013 – IOS Press and the authors. All rights reserved

21

AUTHOR COPY

In this paper, a model has been developed that in-

corporates biases in a model for trust dynamics. In

order to do so, an existing trust model is taken as a

point of departure (cf. [11]), which was applied, for

example, in [17–19]). Biases have been added to this

model using a number of different approaches for the

manner in which biases affect the level of trust. In-

troducing a trust model with the purpose to model

human behaviour in a more realistic way requires a

thorough evaluation of the model. Therefore in this

paper, a number of approaches have been used to

evaluate the introduced models. First of all, the be-

haviour of the models themselves have been rigor-

ously compared and analyzed using identified emerg-

ing properties. Also, an extensive mathematical

analysis of monotonicity, equilibria and behaviour

around equilibria has been performed for this purpose.

In addition to these types of formal analysis, also an

empirical analysis has been performed. The models

have been validated against empirical data that has

been obtained from an experiment conducted with

human subjects. Such a full empirical validation is

not so common for computational trust models.

However, some authors have done some form of

validation. For instance, in (Jonker, Schalken,

Theeuwes and Treur 2004) an experiment has been

conducted whereby the trends in human trust behav-

iour have been analyzed to verify properties underly-

ing trust models developed in the domain of multi-

agent systems. However, no attempt was made to

exactly fit the model to the trusting behaviour of the

human. The outcome of the validation experiment

presented in the current paper shows that the intro-

duced bias-based models perform significantly better

than comparable models without explicit representa-

tion of biases.

This paper is organized as follows. First, in Sec-

tion 2 six new human bias-based trust models are

introduced across computational and human cogni-

tive dimensions. Thereafter, simulation results of

these bias-based trust models are presented in Sec-

tion 3. The formal analyses of the newly designed

bias-based trust models through logical and mathe-

matical means are described in Sections 4 and 5,

respectively. Thereafter, the human-based trust ex-

periment is explained in Section 6. The validation

results of the models based on empirical data col-

lected in the experiment described in Section 6 are

presented in Section 7, and finally, Section 8 is a

discussion.

2. Models for biased trust dynamics

In this section a number of trust models are pro-

posed that incorporate biased human behaviour. In

order to be able to model bias-based trust dynamics,

an existing trust model aimed at representing human

trust is taken as a basis. This is a well-known model

presented in [11] and applied, for example, in [17–

19]. The model is expressed as follows:

∆

∆ 1

In this trust model, it is assumed that the human re-

ceives a certain experience at each time point, E(t).

The experience is represented by a value in the inter-

val [0, 1]. It is then compared with the current trust

level T(t) and the difference is multiplied with a trust

update speed factor γ. This difference is then multi-

plied by the chosen step size ∆t and added to the

current trust level to obtain a new trust level.

The model described above does not include bi-

ases; therefore in this paper extensions of the model

are introduced incorporating biases. This can be done

in different manners. It is assumed that human biases

can affect trust in a number of ways. More specifi-

cally, there are different ways in which the bias plays

a role in the formation of a new trust value; this is

referred to as the cognitive dimension in Fig. 1. In

this paper, three options are distinguished:

(a) the bias solely plays a role in the way in which

the human perceives an experience with a spe-

cific trustee: the experience is transformed

from a certain objective value to a subjective

biased experience value, which is then used to

derive a new trust value,

(b) the experience is again perceived differently

based upon the bias, but the current trust value

also plays a role in the perception of the ex-

perience,

(c) the experiences are not biased, but the trust

value itself is biased.

Besides these different possibilities of modelling the

point at which the bias plays a role in the trust forma-

tion process, the precise way in which the bias is

incorporated within the model can also be varied.

There can be assumed a more linear trend in the bias

behaviour, or a logistic type of trend can be assumed;

this is referred to as the computational dimension in

Fig. 1. Given these dimensions, in total 6 models for

incorporating trust in the unbiased model expressed

in Eq. (1) can now be formulated (see Fig. 1):

M. Hoogendoorn et al. / Modelling biased human trust dynamics22

AUTHOR COPY

1. linear model with biased experience,

2. linear model with biased experience influenced

by current trust,

3. linear model with bias solely determined by

current trust,

4. logistic model with biased experience,

5. logistic model with biased experience influ-

enced by current trust,

6. logistic model with bias solely determined by

current trust.

The above models are abbreviated as LiE, LiET, LiT,

LoE, LoET, and LoT respectively. In order to incor-

porate the biased behaviour in the model presented in

Eq. (1), functions have been defined that take the

current experience (for models LiE and LoE), the

experience and the trust (for models LiET and LoET),

or the trust value itself (for models LiT and LoT) and

transforms that into a biased value. This biased value

can then be used to calculate the new trust value

based upon Eq. (1).

2.1. Trust models with biased experience

For the models that express the bias solely based

upon the experience, the following two equations are

used (for linear and logistic respectively):

LiE:

2 11 0.5

2 0.5

LoE:

1 1൫ିఙሺாሺ௧ሻିఛሻ൯

⁄

In the first equation, β is the bias parameter from the

interval [0,1]. Here values for β of 0.0, 0.5 and 1.0

represent an absolute negative, neutral and absolute

positive bias, respectively. It can be seen that for the

case of a positive bias (i.e., β > 0.5) the current ex-

perience is increased with a factor dependent on the

positiveness of the bias (the more positive the bias,

the more the objective experience is increased). For

the logistic equation (LoE), σ and τ are the steepness

and threshold parameters for the logistic transforma-

tion. In the logistic transformation τ is assumed to

represent the human’s bias. It is assumed that this

value has an inverse relationship with β (i.e., τ =

1 – β). Furthermore E(t), and T(t) are the experience

and human trust level on the given trustee at time

point t, respectively. The resulting value of the func-

tion f(E(t)) is the biased experience.

This function can be incorporated into the base

model (Eq. (1)) in a general setting as follows:

∆

∆ 2

For the specific (linear and logistic) cases considered

this becomes:

∆

2 11

∆ 0.5

∆

2∆ 0.5

∆

1 1൫ିఙሺாሺ௧ሻିఛሻ൯

⁄∆

2.2. Trust models with biased experience affected by

current trust

In the second set of bias equations, the bias plays a

role in combination with the current trust value and

the experience, as expressed below.

LiET:

,1 11

1

LoET:

,1 1൫ିఙሺாሺ௧ሻା்ሺ௧ሻିఛሻ൯

⁄

The first equation (linear model) expresses that the

more positive the bias is, the more the evaluation will

be increased depending on the distance of the experi-

ence and the trust to the highest value. The second is

the logistic variant of the model, whereby the combi-

nation of the experience and the trust are used in the

threshold function.

The function can be inserted into the base model in

a general setting as follows:

∆

, ∆ 3

For the specific (linear and logistic) cases considered

this becomes:

Fig. 1. Bias-based trust models.

M. Hoogendoorn et al. / Modelling biased human trust dynamics 23

AUTHOR COPY

∆

"1 11

1#∆

∆

1 1൫ିఙሺாሺ௧ሻା்ሺ௧ሻିఛሻ൯

⁄

∆

2.3. Trust models with bias solely determined by

current trust

The final set of equations concerns the bias solely

based upon the trust level, and not on the experience

itself. The following two equations are used for this

purpose:

LiT:

1121

0.5

112

0.5

LoT:

1 1 1൫ିఙሺ்ሺ௧ሻିఛሻ൯

⁄

The equations follow the same structure as seen for

the experience-based bias, except that now the trust

value is used.

For the general setting it is combined with the base

model as follows:

∆

∆ 4

For the specific (linear and logistic) cases consid-

ered this becomes:

∆

"1121 #∆

0.5

∆

" 112#∆

0.5

∆

" 1

1 1൫ିఙሺ்ሺ௧ሻିఛሻ൯

⁄#∆

3. Simulation results for the biased human trust

models

In order to observe the behaviour of bias-based

trust models described in the previous section, sev-

eral simulation experiments are performed. In these

simulation experiments first each model is simulated

independently against a set of experience values and

then these models are compared using a novel tech-

nique called mutual mirroring of models as described

in [9].

3.1. Single model comparisons

In this first experiment, merely one trustee for

which an agent has to form trust is considered. In this

section the results of one of these experiments is

presented in detail. In Table 1 the experimental con-

figuration for this simulation is described. Here it can

be seen that bias parameter is changed from 0.0 to,

0.5 and 1.0 which represents negative, neutral and

positive bias respectively. For comparison purposes,

the bias parameter τ for the logistic model is calcu-

lated by means of the following equation: τ = 1 – β.

The trust rate change γ is taken as 0.25. Furthermore,

the initial trust value is taken as 0.50 which means

that the human has neutral trust at time point 0. The

step size (Δt) is set to 0.50.

The experience sequence used in this experiment

is represented in Fig. 2. It can be seen that experience

provided in this experiment change periodically be-

tween the values 0.0, 0.5 and 1.0 respectively with a

period of 10 time steps. Each of these experience

values represents negative, neutral and positive ex-

perience respectively. This experience sequence is

used to see the behaviour of these models on and

between varying extremes.

In Figs 3–5 the results of the simulations given the

experience sequence introduced above are shown.

In Fig. 3 the agent has a negative bias towards the

trustee. A simulation for a neutral bias is shown in

Fig. 4, whereas a positive bias is used in Fig. 5. It can

be observed in the case of the negative bias that both

the LiE and LiET converge to no trust (value 0) de-

spite the fact that the trustee gives some positive

Table 1

Experimental configuration for simulation experiments

Quantity Symbol Value

Bias parameter

β

(linear model)

τ (logistic model)

0.00, 0.50, 1.00

1.00, 0.50, 0.00

Trust change rate

Γ

0.25

Time step ∆t 0.50

Initial trust T(0) 0.50

Steepness

Σ

5

Experiences

E

(t) Periodic (0.0, 0.5, 1.0)

on 10 time steps each

M. Hoogendoorn et al. / Modelling biased human trust dynamics24

AUTHOR COPY

experiences. The LiT, LoT, and LoE variants show

almost similar trends compared to the base trust

model but with a much lower trust value (which is

precisely as desired due to the negative bias). The

final variant of the model (LoET) shows an undesired

result: the trust is actually higher than the base model.

This is due to higher parameter value of parameter σ

(steepness) which is 5. For lower values of the steep-

ness (<3) this model shows desired results as well

(but has not been shown for the sake of brevity).

In Fig. 4 a neutral bias, i.e., (β = 0.5 and τ = 0.5,

σ = 5) is used, and all the models except for one

show behaviour similar to the baseline model (which

is as expected as there is no bias). The LoET does

however show very different and undesirable behav-

iour as it converges to maximum trust value. This

relates to the fact that for this type of model the value

0.5 does not show an upward-downward symmetry

as required for a non-biased case. Therefore this

model does not qualify well in this respect.

In Fig. 5 an absolute positive bias is set (i.e., β = 1

and τ = 0, σ = 5). In the figure, the LiE. LiET, and

LoET converge to maximum trust (value 1) despite

the fact that the trustee gives some negative experi-

ences. This behavior is not completely as desired, but

could be adjusted by taking a different steepness

value. LoE, LiT and LoT show an almost similar

trend as the baseline trust model does, but with

higher in trust value, precisely is as desired.

3.2. Mutual mirroring of the bias-based trust models

To analyze the generalization capacity of these

models a novel technique named mutual mirroring of

models is used as introduced in [9]; see also [7]. In

this method, a specific trace (simulation run) of a

source model is taken as a basis, and a parameter

tuning approach (e.g., exhaustive search within the

parameter space) for a target model is performed to

see how closely the target model can describe the

trace of the source model (i.e., what the set of pa-

rameters is with minimum error). This gives a good

indication how much the models can describe each

others’ behaviour, and some indication of similarity.

The mirroring is also done in the opposite direction

(i.e., from a trace of the target model to parameters of

the source model). This process of mirroring both

Fig. 3. Simulation results for absolute negative bias (β = 0 and

τ = 1, σ = 5).

Fig. 2. Experience sequence.

Fig. 4. Simulation results for neutral or no bias (β = 0.5 and

τ = 0.5, σ = 5).

Fig. 5. Simulation results for absolute positive bias (β = 1 and

τ = 0, σ = 5).

020 40 60 80 100 120 140 160 180

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Time Step

Trust Value

UM

LiE

LiET

LiT

LoE

LoET

LoT

10 20 30 40 50 60 70 80 90 10011012013014 0150160170180

0

0.25

0.5

0.75

1

Time Step

Experience Value

020 40 60 80 100 120 140 160 180

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Time Step

Trust Value

UM

LiE

LiET

LiT

LoE

LoET

LoT

020 40 60 80 100 120 140 160 180

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Time Ste p

Trust Value

UM

LiE

LiET

LiT

LoE

LoET

LoT

M. Hoogendoorn et al. / Modelling biased human trust dynamics 25

AUTHOR COPY

models into each other is called mutual mirroring of

models. The mirroring process can provide a good

indication on the similarity of models. For more

details on the approach see [7,9].

The mirroring techniques have been applied to the

models introduced in Section 2. The results are

shown in Table 2. Here, the columns represent the

target models while the rows represent the source

models.

For a specific trace of the source model (given a

certain set of parameter settings) the parameters of

the target model are exhaustively searched to gener-

ate behaviour similar to the trace of the source model

with minimum root mean squared error. The values

in each cell of the table represent the average error

for nine different source model traces generated with

different bias values and experience sequences. In the

first row of the table it can be seen that on average

the source model LiE can be approximated using the

LiE, LiET, LiT, LoE, LoET and LoT variants with

error of 0.00, 0.04, 0.22, 0.12, 0.14 and 0.22 respec-

tively. Furthermore in the last column of the first row

it can be seen that the average error of the mirroring

process with all other models is 0.12. This seems to

be the most difficult behaviour to approximate on

average as the other rows show a lower average value.

Especially the behaviour of the LiT and LoE can be

very well approximated by the other models. Fur-

thermore, in the last row the values are shown that

indicate how well a model can describe the other

model’s behaviour. This shows that LiE and LiET

can describe many of the source models very well.

4. Logical verification of the bias-based trust

models

When developing a new model, a thorough analy-

sis of the behaviour is required to have sufficient

confidence in the appropriate behaviour of the model.

One way to perform such an analysis is to conduct a

mathematical analysis (see Section 5). However,

given the complexity of the models proposed in this

paper, the analysis of more complex (temporal) pat-

terns might not be feasible using these techniques.

Therefore, in this section, certain desired emergent

properties are discussed with respect to the bias-

based trust models that express complex patterns

over time. To show that the models indeed generate

this desired behaviour, these properties have been

verified upon the simulation traces that have been

produced by the models proposed in Section 2. This

does not prove a complete adherence of the model to

the properties, but it does shown that for the selected

simulation runs (which are of course carefully se-

lected in order to have representative results) adhere

to the properties or not. In order to perform this veri-

fication in an automated fashion, the hybrid temporal

language TTL (Temporal Trace Language, cf. [2,

16]) and its software environment has been used. In

addition to a dedicated editor TTL features an auto-

mated verification tool that automatically verifies

specified properties against traces that have been

loaded in the verification tool. The language TTL is

explained first, followed by a presentation of the

desired properties related to trust.

4.1. Temporal trace language (TTL)

The hybrid temporal language TTL supports for-

mal specification and analysis of dynamic properties,

covering both qualitative and quantitative aspects.

TTL is built on atoms referring to states of the world,

time points and traces, i.e., trajectories of states over

time. In addition, dynamic properties are temporal

statements that can be formulated with respect to

traces based on the state ontology Ont in the follow-

ing manner. Given a trace γ over state ontology Ont,

the state in γ at time point t is denoted by state(γ, t).

These states can be related to state properties via the

formally defined satisfaction relation denoted by the

infix predicate |=, i.e., state(γ, t) |= p denotes that

state property p holds in trace γ at time t. Based on

these statements, dynamic properties can be formu-

lated in a formal manner in a sorted first-order predi-

cate logic, using quantifiers over time and traces and

the usual first-order logical connectives such as ¬, ∧,

∨, ⇒, ∀, ∃. As a built-in construct in TTL, summa-

tions can be expressed, indexed by elements X of a

sort S:

∑X:S case(ϕ(X), V1, V2)

Table 2

Results for mutual mirroring of the models

Source

model

Target model

LiE LiET LiT LoE LoET LoT AVG

LiE 0.00 0.04 0.22 0.12 0.14 0.22 0.12

LiET 0.02 0.00 0.19 0.10 0.13 0.19 0.11

LiT 0.01 0.03 0.00 0.01 0.06 0.00 0.02

LoE 0.01 0.03 0.09 0.00 0.08 0.09 0.05

LoET 0.03 0.05 0.23 0.11 0.00 0.22 0.11

LoT 0.01 0.02 0.00 0.01 0.05 0.00 0.02

AVG 0.02 0.03 0.12 0.06 0.08 0.12

M. Hoogendoorn et al. / Modelling biased human trust dynamics26

AUTHOR COPY

Here for any formula ϕ(X), the expression

case(ϕ(X), V1, V2)

indicates the value V1 if ϕ(X) is true, and V2 other-

wise. For example,

∑X:S case(ϕ(X), 1, 0)

simply denotes the number of elements X in S for

which ϕ(X) is true. As expressing counting and sum-

mation in a logical format in an elementary manner

in general leads to rather complex formulae, this

built-in construct is very convenient in use. For more

details on TTL and the precise functioning of the

checker tool, see [2,16].

4.2. Verification of bias-based trust models

This section describes verification process for the

bias-based trust models presented in Section 2. First,

in Section 4.2.1 the properties that have been identi-

fied for bias-based trust models are introduced and

then in Section 4.2.2 results of the checks are pre-

sented.

4.2.1. Properties for bias-based trust models

Four properties have been identified with respect

to biased behaviour of human trust. The first property

expresses the general principle of the bias, namely

that once a person has a more positive bias towards a

trustee, this trustee will more frequently be the most

trusted trustee, as expressed in property P1 below.

Note that in this property (and also for properties P2

and P3), it is assumed that the bias does not change

during the simulation, and hence, the value at the first

time point is selected.

P1: General bias property. If within two traces

with the same experience sequence in one trace an

agent has a more positive bias towards a trustee com-

pared to the other trace, and the agent has the same

biases for the other trustees, then the trustee will

more frequently be the trustee with the highest trust

value in the trace with the higher bias compared to

the trace with the lower bias. For example, this then

results in this trustee being selected more frequently.

The formalization of the property is shown below.

First, it is checked whether the traces that are being

compared contain the same experience sequence.

Furthermore, it is checked whether the biases for the

trustee tr1 considered different (and in fact, is higher

in the first trace). Note that this comparison is done at

time point 0 as it is assumed that the bias does not

change over time in a single run. Furthermore, it is

checked whether there exists a single bias value the

agents has for all other trustees in both traces, then

one sums the cases where the trustee tr1 is the trustee

with the highest trust value and this amount should

be higher in the first trace compared to the second.

P1 ≡ ∀γ1, γ2:TRACE, tr1:TRUSTEE, b1, b2:REAL

[ same_experience_sequence(γ1, γ2) &

state(γ1, 0) |= bias_for_trustee(tr1, b1) &

state(γ2, 0) |= bias_for_trustee(tr1, b2) & b1 > b2 &

∀tr2:TRUSTEE ≠ tr1 ∃b3:REAL

[state(γ1, 0) |= bias_for_trustee(tr2, b3) &

state(γ2, 0) |= bias_for_trustee(tr2, b3) ] ⇒

[ ∑t:TIME case(highest_trust_value(γ1, t, tr1), 1, 0) ≥

∑t:TIME case(highest_trust_value(γ2, t, tr1), 1, 0) ] ]

Here the same experience sequence is simply a prop-

erty expressing that the experience values in both

traces should be the same:

same_experience_sequence(γ1:TRACE, γ2:TRACE,) ≡

∀t:TIME, tr:TRUSTEE, v:REAL

[ state(γ1, t) |= objective_experience_value(tr, v) ⇒

state(γ2, t) |= objective_experience_value(tr, v) ]

In the formalisation of the predicate indicating the

highest trust value which is used in P1 the trust value

for the trustee considered is bound by the ∀–

quantifier. For this value it is then checked whether

for all other trustees and trust values encountered no

higher value than the value for trustee tr1 is encoun-

tered.

highest_trust_value(γ:TRACE, t:TIME, tr1:TRUSTEE) ≡

∀v1:REAL

[ state(γ, t) |= trust_value(tr1, v1) ⇒

∀tr2:TRUSTEE ≠ tr1, v2:REAL

[ state(γ, t) |= trust_value(tr2, v2) ⇒ v2 < v1 ] ]

The second property expresses that the trust level

itself will be higher in the case of a more positive

bias.

P2: Trust comparison. Trustees for which an agent

has a more positive bias have a higher trust value

compared to a trace in which the agent has a lower

bias with respect to the trustee (given that the experi-

ences are equal as well as the biases for the other

trustees).

The formalization of this property is very similar

to P1, except that now a comparison is made between

the trust values themselves.

M. Hoogendoorn et al. / Modelling biased human trust dynamics 27

AUTHOR COPY

P2 ≡ ∀γ1, γ2:TRACE, tr1:TRUSTEE, b1, b2:REAL

[ same_experience_sequence(γ1, γ2) &

state(γ1, 0) |= bias_for_trustee(tr1, b1) &

state(γ2, 0) |= bias_for_trustee(tr1, b2) & b1 > b2 &

∀tr2:TRUSTEE ≠ tr1 ∃b3:REAL

[ state(γ1, 0) |= bias_for_trustee(tr2, b3) &

state(γ2, 0) |= bias_for_trustee(tr2, b3) ] ⇒

∀t:TIME, tv1, tv2:REAL

[ state(γ1, t) |= trust_value(tr1, tv1) &

state(γ2, t) |= trust_value(tr1, tv2) ] ⇒ tv1 ≥ tv2 ]

In order to facilitate the addition of bias to existing

models, a translation scheme has been proposed to

translate objective experiences into subjective ex-

periences (i.e., experiences coloured by the bias). In

case of a more positive bias, the biased experiences

will be at least as high.

P3: Experience comparison. The objective experi-

ence provided by a trustee is translated into a higher

subjective experience for trustees for which the agent

has a higher bias (given the same experience se-

quence).

The formalization of this property takes the first

part which is by now well-known from P1 and P2 as

an antecedent and checks to see whether the subjec-

tive experiences are indeed at least as high for the

trace in which a higher bias is encountered.

P3 ≡ ∀γ1, γ2:TRACE, tr:TRUSTEE, b1, b2:REAL

[ [ same_experience_sequence(γ1, γ2) &

state(γ1, 0) |= bias_for_trustee(tr, b1) &

state(γ2, 0) |= bias_for_trustee(tr, b2) & b1 > b2] ⇒

∀t:TIME, ev1, ev2:REAL

[ [ state(γ1, t) |= subjective_experience_value(tr, ev1) &

state(γ2, t) |= subjective_experience_value(tr, ev2) ]

⇒ ev1 ≥ ev2 ] ]

Finally, in some of the bias model, trust is explic-

itly considered to colour the experiences. In case the

trust level is higher, the same objective experience

gets an even more positive value.

P4: Influence of trust upon experience. If the trust

level for a certain trustee at time point t is higher than

the trust level at another time point t’, whereas the

objective experience is equal and not on the bound-

ary of the scale (i.e., 0 or 1), then the subjective ex-

perience will be higher at time point t.

The formalization of this property is a bit more

complicated. First, the property binds the trust value

at a time point t for a certain trustee as well as the

objective experience. Hereby, a check is performed

to make sure the objective experience is neither 0 nor

1 as this would sometimes make it impossible to have

a higher subjective value. Given that this is the case,

and given that the objective experience is the same at

another time point t’ at which the trust value is lower

compared to the trust value at time t, this means that

the subjective value at time t must be higher.

P4 ≡ ∀γ:TRACE, t, t’:TIME, tr:TRUSTEE,

tv1, tv2, ov, sv1, sv2:REAL

[ state(γ, t) |= trust_value(tr, tv1) &

state(γ, t) |= objective_experience_value(tr, ov) &

ov > 0 & ov < 1 &

state(γ, t) |= subjective_experience_value(tr, sv1) &

state(γ, t’) |= trust_value(tr, tv2) & tv1 > tv2 &

state(γ, t’) |= objective_experience_value(tr, ov) &

state(γ, t’) |= subjective_experience_value(tr, sv2) ]

⇒ sv1 > sv2

4.2.2. Verification results for bias-based trust models

Based upon the traces resulting from simulations

of the trust models so-called traces have been gener-

ated. These traces are essentially logs of the simula-

tions that indicate for each time point what states

hold. These traces are loaded into the TTL Checker

software which then expresses whether a property

(i.e., P1–P4) holds for the trace (or a combination of

traces) or not. The results of the verification are

shown in Table 3. It can be seen that property P1 is

satisfied for all bias models presented in this paper.

When looking at the properties P2 and P3 however,

the properties also hold for the various models that

have been identified. Finally, property P4 is only

satisfied for the models where trust is considered

when forming the subjective experience, which

makes sense as this property precisely, describes this

influence. Properties P3 and P4 are actually not rele-

vant for models LoET and LoT as they do not incor-

porate the notion of subjective experience, therefore

the property is always satisfied (due to the fact that

the antecedent of the implication never holds).

5. Mathematical analysis of bias-based trust

models

The models explored in this paper are adaptive

with respect to the experiences of the agent. This

means, for example, that when in a time period with

very positive experiences, also trust will reach higher

M. Hoogendoorn et al. / Modelling biased human trust dynamics28

AUTHOR COPY

levels, and in periods with less positive experiences

trust levels will go down. For very long periods of

experiences of the same level, the trust level will

reach some stable level, which is an equilibrium for

the model for the given experience level. It gives a

more deepened insight in the model when it is known

what the value of such an equilibrium is for a given

experience level: the model will drive the trust level

in the direction of that value. Moreover, the speed by

which such a convergence process takes place also is

useful information about a model. For these types of

analyses the techniques used in the previous section

are not practical to use, but mathematical techniques

are available that can be used quite well.

The properties addressed here by such mathemati-

cal techniques focus for a given point in time t in

particular on criteria that determine whether due to a

given experience the trust level will increase, de-

crease or will be in equilibrium. Moreover for the

equilibria of the models, the behaviour near such

equilibria is addressed: whether they are attracting or

not, and how fast the convergence takes place. These

properties are much more specific and limited com-

pared to the wider types of properties addressed in

Section 4, but the mathematical methods allow for

more in depth results.

First the general case is addressed; in Table 4 an

overview of the results for the general case is sum-

marised. Next, the analysis is made more specific for

the case of linear functions; at the end of the section

in Table 5 an overview of the results for these spe-

cific linear functions is presented. Note that the

analysis is done for any given time point t, which is

sometimes indicated as an argument, but will some-

times be left out to get expressions more transparent.

5.1. Mathematical analysis of trust models with

biased experience

Recall that for the models that express the bias

solely based upon the experience, the following dif-

ference equation is used.

where it is assumed that γ > 0. Note that from the

equation above it immediately follows:

So, in this case the following criteria can be obtained

for trust models with biased experiences:

Equilibrium, increasing and decreasing: trust models

with biased experiences

(a) T is in equilibrium for a given if and only if

,

(b) T is increasing if and only if ,

(c) T is decreasing if and only if .

For example, (b) shows a criterion for an experi-

ence to let the trust level increase. If the trust already

has some level T, it can only increase when an ex-

perience with level E at least satisfying

occurs; otherwise trust will decrease or stay the same.

Another way to use this is to determine directly to

which equilibrium trust can go if a given experience

level E is constantly offered; according to criterion

(a) this equilibrium level for trust is . Further-

more, from the monotonicity criteria above it can be

derived in the following manner that the equilibrium

is always attracting. Suppose Teq is an equilibrium for

E, and

; this implies

and therefore T is increasing for the given E by the

criterion (b) above. Similarly, when

for the

given E it is found that T is decreasing by crite-

rion (c). This proves that the process will always

converge to the equilibrium, independent of the func-

tion . This will also be confirmed by the analysis of

the behaviour around the equilibrium below.

Determining the behaviour around an equilibrium

Independent of the precise form of the function

(and hence also independent of the bias parameter β),

the behaviour around an equilibrium for a given

constant experience can be found here as follows.

Write

δ, with δ the deviation of

T from the equilibrium Teq for which it holds

.

Table 3

Result of verification

LiE LiET LiT LoE LoET LoT

P1 satisfied satisfied satisfied satisfied satisfied satisfied

P2 satisfied satisfied satisfied satisfied satisfied satisfied

P3 satisfied satisfied satisfied satisfied satisfied satisfied

P4 failed satisfied failed satisfied satisfied satisfied

M. Hoogendoorn et al. / Modelling biased human trust dynamics 29

AUTHOR COPY

δ

δ

δ

δδ δ

δδδ

δ

δ

As a differential equation this can be solved analyti-

cally using an exponential function:

δδγ

This shows that the speed of convergence directly

relates to parameter γ, and the convergence rate de-

fined as reduction factor of the deviation per time

unit is

γ

This is independent of β, or the function . More

specifically, since γ > 0, the convergence rate is al-

ways <1; from this it follows that the equilibrium is

always attracting.

This shows that the speed by which trust adapts to

a certain experience level is independent of the spe-

cific function and bias parameter β; it is higher

when γ is higher and lower when γ is lower.

5.2. Mathematical analysis of trust models with

biased experience also affected by trust

For the models that express the bias based both

upon the experience and the current trust level, the

following difference equation was used:

with γ > 0. In a similar manner as above the follow-

ing criteria are obtained:

Equilibrium, increasing and decreasing: Biased ex-

perience also affected by trust

(a) is in equilibrium for a given if and only if

,0

,

(b) is increasing if and only if ,0,

(c) is decreasing if and only if ,0.

This again shows a criterion, for example, for an

experience to let the trust level increase. If the trust

already has some level T, it can only increase when

an experience with level E at time t at least satisfying

,0 is obtained; otherwise trust will decrease

or stay the same.

Furthermore, some criterion on the function can

be found in order that the equilibrium Teq for E is

attracting. Attracting means that if T is close to Teq

with T < Teq, then for the given E it should be the

case that T increases, which according to the above is

equivalent with , > 0. So, starting from T = Teq

with ,

= 0, when T is taken lower, the value

of ,

has to become higher:

⇒

This is equivalent with the criterion that in ,

the function

is decreasing in its second argument:

,

⁄0. Below this will be confirmed

from the analysis of the behaviour around an equilib-

rium. This shows that not all functions will provide

the property that the trust levels converge to such an

equilibrium value. For a choice to be made for some

function this has to be considered. Below it will be

shown that for the choices made in the current paper

this criterion is always fulfilled.

Determining the behaviour around an equilibrium

Depending on the form of the function and also on

the bias parameter β, the behaviour around an equi-

librium for a given constant experience can be

found as follows. Write

δ, with δ(t)

the deviation from the equilibrium for which it

holds ,

0. For the first-order Taylor

approximation around in its second argument is

used, where / denotes the partial derivative of

with respect to its second argument :

Using this it holds

δ

δ

δ

δ

Then the following is obtained:

δ δ⋅

δ

δδ

δ

δ

δ

As a differential equation this can be solved analyti-

cally using an exponential function:

δδ ,

⁄

The convergence rate is defined as reduction

factor of the deviation per time unit; this is

M. Hoogendoorn et al. / Modelling biased human trust dynamics30

AUTHOR COPY

.This provides a condition on

when an equilibrium is attracting, namely

,

⁄0. Note that in this case the con-

vergence speed does not only depend on γ but also

on f, which in principle relates to the bias β. This

speed is higher when γ is higher, but also when

,

⁄ is more negative.

5.3. Mathematical analysis of trust models with bias

solely determined by current trust

For the models that express the bias based only

upon the current trust level, the following difference

equation was used:

where γ > 0. Similarly the following criteria are

found:

Equilibrium, increasing and decreasing: Bias solely

determined by current trust

(a) is in equilibrium for a given if and only if

,

(b) is increasing if and only if ,

(c) is decreasing if and only if .

Like before, this shows a criterion, for example, for

an experience to let the trust level increase. If the

trust already has some level T, it can only increase

when an experience with level E at least satisfying

is obtained; otherwise trust will decrease

or stay the same. Moreover, a criterion on the func-

tion can be found in order that the equilibrium

for E is attracting. As before note that attracting

means that if T is close to with

, then for the

given E it should be the case that T increases, which

according to criterion (b) above is equivalent with

. So, starting from

with

, when T is taken lower, the value of

becomes lower:

⇒

This means that in the function has to be in-

creasing:

⁄0. Below, this criterion for

being attracting will be confirmed when the behav-

iour around an equilibrium is analysed. This shows

again that not all functions will provide the prop-

erty that the trust levels converge to an equilibrium

value. For a choice to be made for some function

this criterion

⁄0 has to be taken into

account. Below it will be shown that for the choices

made in the current paper this criterion is always

fulfilled.

Determining the behaviour around an equilibrium

Again, depending on the form of the function and

also on the bias parameter β, the behaviour around an

equilibrium for a given constant experience can be

found as follows. Write

δ, with δ

the deviation from the equilibrium for which it

holds

. For the first-order Taylor ap-

proximation around is used:

Using this it is obtained:

δ

δ

δ

δδ δ

δδ

δ

δδ

δ

δ

δ

As a differential equation this can be solved analyti-

cally using an exponential function:

δδ

⁄

This shows that the speed of convergence does not

only relate to parameter γ, but also to /

which in principle relates to the bias β. The conver-

gence rate defined as reduction factor of the devia-

tion per time unit is

⁄

So, also in this case the convergence speed does not

only depend on γ but also on , which in principle

relates to the bias β. This speed is higher when γ is

higher, but also when / is higher.

5.4. Mathematical analysis of the example biased

trust models for the three types

In this section, for each of the three general types

of biased trust models analysed above, it will be

investigated how the criteria can be formulated more

specifically for the linear functions used in the cur-

M. Hoogendoorn et al. / Modelling biased human trust dynamics 31

AUTHOR COPY

rent paper as instances for the function : LiE, LiET,

and LiT.

5.4.1. More specific analysis for the linear case of

bias only depending on experience (LiE)

For the first case the following linear function was

addressed (LiE):

Case β ≥ 0.5

Criterion for increasing for LiE with β ≥ 0.5

β

β

β

β

β

β

½ β

Criterion for decreasing for LiE with β ≥ 0.5

½ β

Criterion for equilibrium for LiE with β ≥ 0.5

½ β

ββ½

½β β

β β

β

Note that for β = 0.5 (no bias) the criterion for an

equilibrium is = , what is to be expected. For β =

0.75, the criterion is

½

Note that for lower values of T this can provide a

negative number. However, as the experience cannot

be lower than 0, this implies that for such values of

no equilibrium occurs. For β = 0.875, the criterion is

½

For β approaching 1, the criterion always becomes a

negative number (implying increase), unless = 1;

this implies that for this value of β no equilibrium

occurs except for = 1 and any value for .

Behaviour around the equilibrium for LiE with

β ≥ 0.5 For this case the behaviour around the equi-

librium does not depend on the specific form of the

function . The convergence rate is: γ, which

is independent, for example, of β. As γ > 0, the equi-

librium is always attracting.

Case β ≤ 0.5

Criterion for increasing for LiE with 0.5

½

Criterion for decreasing for LiE with 0.5

½

Criterion for equilibrium for LiE with 0.5

½

Behaviour around the equilibrium for LiE with

0.5 For this case the behaviour around the equi-

librium does not depend on the specific form of the

function . The convergence rate is: γ, which

is independent, for example, of β or E. As γ > 0, the

equilibrium is always attracting.

5.4.2. More specific analysis for the linear case of

bias depending on both experience and trust

(LiET)

For the second case the following linear function

was addressed (LiET):

For the linear example the inequalities and equa-

tion can be explicitly solved as follows.

Table 4

Results of the mathematical analysis for the general case

Bias depends on Increasing/decreasing Equilibrium value Convergence rate Attracting

only on experience

,

γ

always

on experience and trust

, 0, , 0 ,

0 ,

⁄

,

⁄< 0

only on trust ,

⁄

⁄>0

M. Hoogendoorn et al. / Modelling biased human trust dynamics32

AUTHOR COPY

Criterion for increasing for LiET

β

ββ

ββ

β β

Criterion for decreasing for LiET

β β

Criterion for equilibrium for LiET

β β

β

β β

ββ

ββ

ββ

Behaviour around the equilibrium for LiET For the

specific linear function used above, it holds:

Using this, for the linear case it is obtained:

δδ

and the convergence rate is . This

shows that for this case the speed of convergence not

only relates to parameter γ, but also to β and . More

specifically, the convergence rate is < 1 if and only if

This is a condition for an equilibrium to be attracting.

It can be rewritten into an explicit criterion for E as

follows:

This is always the case.

5.4.3. More specific analysis for the case of bias

depending only on trust (LiT)

For the third case the following function was ad-

dressed (LiT):

when β ≥ 0.5

when β ≤ 0.5

This can be analysed more specifically as follows

Case β ≥ 0.5

Criterion for increasing for LiT with β ≥ 0.5

Criterion for decreasing for LiT with β ≥ 0.5

Criterion for equilibrium for LiT with β ≥ 0.5

For the special case that β = 0.5 (no bias) this latter

criterion reduces to a linear equation – + = 0

with solution = . For the general case β > 0.5 the

above expression is a quadratic equation in with

discriminant

βββ

βββββ

βββββ

ββ

ββ

β

From this expression for , which is linear in both β

and , given that β ≥ 0.5 it can easily be seen that

is always ≥ 1:

• for β = 0.5 it holds 4

1

431,

• for β = 1 it holds

8

1

43

541

since ≤ 1.

Alternatively, considering special values of :

• for = 1 it holds = 1,

• for = 0 it holds 8β3!431

since β ≥ 0.5.

Therefore is positive and the quadratic equation

has two solutions for T

,

β

β

ββ

β

Since ≥ 1 for the highest solution it holds

β βββ

β β

β

β

Similarly, from ≥ 1 it follows that for the lowest

solution (for the –) it holds

M. Hoogendoorn et al. / Modelling biased human trust dynamics 33

AUTHOR COPY

ଵ 4β 1 1/22 1 4β 2/22 1 1

Therefore the equilibrium for a given is the

lowest solution

= β√

β = β√β

β ≤ 1

This is a positive number since √D ≤ (4β – 1) as can

be seen from the initial expression

ββββ

Behaviour around the equilibrium for LiT with β ≥

0.5 It holds

β

β

Therefore for this case the convergence rate is

/β

This depends both on γ and β, and via also on .

The criterion for the equilibrium being attracting is

that $/$ 0. This is equivalent to:

β

As β ≥ 0.5, this is always the case.

Case β ≤ 0.5

Criterion for increasing for LiT with β ≤ 0.5

β

Criterion for decreasing for LiT with β ≤ 0.5

β

Criterion for equilibrium for LiT with β ≤ 0.5

β

β

This is a quadratic equation in with discriminant

β

Then

, β√

β β

Solutions for require that ≥ 0; this is equivalent

to:

ββ

ββ

ββ

ββ

As ≤ 1 and 1β/12β1, this is always

fulfilled. The highest solution is > 1 as can be

seen from

ββββ

β

β β

β

Therefore the equilibrium value is the smallest

solution

ββββ

As above it can be seen that this is a positive number.

Behaviour around the equilibrium for LiT with β ≤

0.5 It holds

β

β

Therefore for this case the convergence rate is

/ β

This depends both on γ and β, and via also on .

The criterion for the equilibrium being attracting is

that / 0. This is equivalent to:

β

As β ≤ 0.5, this is always the case.

6. Human-based trust experiment

In this section the human-based trust experiment is

explained. In Section 6.1 the participants are de-

scribed. In Section 6.2 an overview of the used ex-

perimental environment is given. Thereafter, the

procedure of the experiment and data collection is

explained in Section 6.3.

6.1. Participants

Eighteen participants (eight male and ten female)

with an average age of 23 (SD = 3.8) participated in

the experiment as paid volunteers. Non-colour

blinded participants were selected. All were experi-

M. Hoogendoorn et al. / Modelling biased human trust dynamics34

AUTHOR COPY

enced computer users, with an average of 16.2 hours

of computer usage each week (SD = 9.32).

6.2. Task

As the bias-based trust models are designed to

work in situations in which humans have to decide to

trust either one of multiple heterogeneous trustees,

the experimental task used involved three different

trustees, namely two human participants and a sup-

port system. The task was a classification task in

which the two participants on two separate personal

computers had to classify geographical areas accord-

ing to specific criteria as areas that either needed to

be attacked, helped or left alone by ground troops.

The participants needed to base their classification on

real-time computer generated video images that re-

sembled video footage of real unmanned aerial vehi-

cles (UAVs). On the camera images, multiple objects

were shown. There were four kinds of objects: civil-

ians, rebels, tanks and cars. The identification of the

number of each of these object types was needed to

perform the classification. Each object type had a

score (either –2, –1, 0, 1 or 2, respectively) and the

total score within an area had be determined. Based

on this total score the participants could classify a

geographical area (i.e., attack when above 2, help

when below –2 or do nothing when in between).

Participants had to classify two areas at the same

time and in total 98 areas had to be classified. Both

participants did the same areas with the same UAV

video footage.

During the time a UAV flew over an area, three

phases occurred: The first phase was the advice

phase. In this phase both participants and a support-

ing software agent gave an advice about the proper

classification (attack, help, or do nothing). This

means that there were three advices at the end of this

phase. It was also possible for the participants to

refrain from giving an advice, but this hardly oc-

curred. The second phase was the reliance phase. In

this phase the advices of both the participants and

that of the supporting software agent were communi-

cated to each participant. Based on these advices the

participants had to indicate which advice, and there-

fore which of the three trustees (self, other or soft-

ware agent), they trusted the most. Participants were

instructed to maximize the number of correct classi-

fications at both phases (i.e., advice and reliance

phase). The third phase was the feedback phase, in

which the correct answer was given to both partici-

pants. Based on this feedback the participants could

update their internal trust models for each trustee

(self, other, software agent).

In Fig. 6 the interface of the task is shown. The

map is divided in 10 × 10 areas. These boxes are the

areas that were classified. The first UAV starts in the

top left corner and the second one left in the middle.

The UAVs fly a predefined route so participants do

not have to pay attention to navigation. The camera

footage of the upper UAV is positioned top right and

the other one bottom right.

The advice of the self, other and the software agent

was communicated via dedicated boxes below the

camera images. The advice to attack, help, or do

nothing was communicated by red, green and yellow,

respectively. On the overview screen on the left,

feedback was communicated by the appearance of a

green tick or a red cross. The reliance decision of the

participant is also shown on the overview screen

Table 5

Results of the mathematical analysis for the specific linear functions

Bias depends on Increasing/decreasing Equilibrium value Convergence rate Attracting

LiE only on experience:

β ≥ 0.5

1½ 1 /1 β

1 ½ 1 /1 β

1½1

/1 β

1 2 1 β1

Always

only on experience:

β ≤ 0.5

½/

½/

½

/

2

Always

LiET on experience and

trust

1β/ 12β

1β/ 12β

1 β

/ 1 2β

/

1β12β

Always

LiT only on trust:

β ≥ 0.5

1 2 11

1 2 11

1

2 11

T

e

q

=β √ β

Always

only on trust:

β ≤ 0.5

2

1β

1 2

2

1β

1 2

2

1β

1 2

T

e

q

= ββ

Always

M. Hoogendoorn et al. / Modelling biased human trust dynamics 35

AUTHOR COPY

behind the feedback (feedback only shown in the

feedback phase). The phase depicted in Fig. 6 was

the reliance phase before the participant indicated his

reliance decision.

6.3. Data collection

During the above described experiment, input and

output were logged using a client-server application.

The interface of this application is shown in Fig. 7.

Two other client machines, that were responsible for

executing the task as described in the previous sub-

section, were able to connect via a local area network

to the server, which was responsible for logging all

data and communication between the clients. The

interface shown in Fig. 7 could be used to set the

client’s IP-addresses and ports, as well as several

experimental settings, such as how to log the data. In

total the experiment lasted approximately 15 minutes

per participant.

Experienced performance feedback of each trustee

and reliance decisions of each participant were

logged in temporal order for later analysis. During

the feedback phase the given feedback was translated

to a penalty of either 0, 0.5 or 1, representing a good,

Fig. 6. Interface of the task.

Fig. 7. Interface of the application used for gathering validation

data (Connect), for parameter adaptation (Tune) and validation of

the trust models (Validate).

M. Hoogendoorn et al. / Modelling biased human trust dynamics36

AUTHOR COPY

neutral or poor experience of performance, respec-

tively. During the reliance phase the reliance deci-

sions were translated to either 0 or 1 for each trustee

Si, which represented that one relied or did not rely

on Si.

7. Validation of bias-based trust models

In this section the validation process of the trust

models described in Section 2 are presented. In Sec-

tion 7.1 the parameter adaption technique is ex-

plained, Sections 7.2 and 7.3 explain the model vali-

dation process and results for the bias-based trust

models, respectively.

7.1. Parameter adaptation

The data collection described in Section 6.3 was

repeated twice on each group of two participants,

called condition 1 and condition 2, respectively. The

data from one of the conditions was used for parame-

ter adaptation purposes for each model, and the data

from the other condition for model validation (see

Section 6.3). This process of parameter adaptation

and validation was balanced over conditions, which

means that conditions 1 and 2 switch roles, so condi-

tion 1 is initially used for parameter adaptation and

condition 2 for model validation, and thereafter con-

dition 2 is used for parameter adaptation and condi-

tion 1 for model validation (i.e., cross-validation).

Then the average was calculated of the two calcu-

lated validities, per participant, per model. This last

value is called the accuracy of the models. The re-

sults are in the form of accuracies per trust model and

their differences are detected using a repeated meas-

ures analysis of variance (ANOVA) and post-hoc

Bonferroni t-tests.

After the different models were tuned, the best fit

model (with the maximum accuracy) is selected

based on the maximum accuracy for the participant at

hand. This was done because at the moment one does

not know beforehand which bias type will be suitable

for the specific participant. The results of the valida-

tion process are in the form of accuracies per trust

model (unbiased model (UM), LiE, LiT, LiET, LoE,

LoT, LoET and the best fit model (MAX)).

Both the parameter adaptation and model valida-

tion procedure was done using the same application

as was used for gathering the empirical data. The

interface shown in Fig. 7 could also be used to alter

validation and adaptation settings, such as the granu-

larity of the adaptation.

The number of parameters of the models presented

in Section 2 to be adapted for each model and each

participant suggest that an exhaustive search as de-

scribed in [6] for the optimal parameter values is

feasible. This means that the entire parameter search

space is explored to find a vector of parameter set-

tings resulting in the maximum accuracy (i.e., the

amount of overlap between the model’s predicted

reliance decisions and the actual human reliance

decisions) for each of the models and each partici-

pant. The corresponding code of the implemented

exhaustive search method is shown in Algorithm 1.

In this algorithm, E(t) is the set of experiences (i.e.,

performance feedback) at time point t for all trustees,

RH(e) is the actual reliance decision the participant

made (on either one of the trustees) given a certain

experience e, RM(e,X) is the predicted reliance deci-

sion of the trust model M, given an experience e and

candidate parameter vector X (reliance on either one

of the trustees), X is the distance between the esti-

mated and actual reliance decisions given a certain

candidate parameter vector X, and δbest is the distance

resulting from the best parameter vector Xbest found

so far. The best parameter vector Xbest is returned

when the algorithm finishes. This parameter adapta-

tion procedure was implemented in Microsoft®

C#.Net 2005 development environment.

In order to compare the different bias-based trust

models described in Section 2, the measurements of

experienced performance feedback were used as

input for the models (i.e., as experiences) and the

output (predicted reliance decisions) of the models

was compared with the actual reliance decisions of

the participant as described in Section 6. It is hereby

assumed that the human always consults the most

trusted trustee. The resulting set of parameters is the

set with minimum error in the prediction of the reli-

ALGORITHM 1: ES-PARAMETER-ADAPTATION(E, RH)

1 δbest= ∞, X = 0

2 for all parameters x in vector X do

3 for all settings of x do

4 δx= 0

5 for all time points t do

6 e = E(t), rM = RM(e, X), rH = RH(e)

7 if rM not equal rH then

8 δx= δx+1

9 end if

10 end for

11 if δx < δbest then

12 Xbest= X, δbest= δx

13 end if

14 end for

15 end for

16 return Xbest

M. Hoogendoorn et al. / Modelling biased human trust dynamics 37

AUTHOR COPY

ance decisions for that specific participant. Hence,

the relative overlap of the predicted and the actual

reliance decisions was a measure for the accuracy of

the models.

7.2. Computational complexity

As the models described in Section 2 have a dif-

ferent number of parameters, the parameter tuning

process took a different amount of time for each of

the models. Assuming that S is the number of sub-

jects, M number of model types (namely unbiased,

linear and logistic), B number of bias types (using

experience, trust, and experience and trust), P the

number of parameters with α degree of precision of

the parameters (in the range of 0–1), T the number of

time steps, and N number of trustees, the complexity

is then O (S.M.B.10Pα.T.N). This indicates that it is

exponential in the number of parameters and their

precision value. The models presented here have a

different number of parameters with different types

of precision. The baseline model has one parameter γ

(with 0.01 precision), while linear models have four

(γ, β1, β2, β3 with 0.01) where β1, β2, and β3 represent

the bias of the subject towards each trustee and the

logistic models have seven parameters (γ, τ1, τ2, τ3, σ1,

σ2, and σ3, where γ and τ has precision 0.01 and σ has

precision 1 within range 1 to 20).

If the time required is, for example, calculated for

LoT for tuning one subject then it has S = 1, M = 1,

B = 1, P = 7 (4 parameters with precision 0.01, and 3

parameters with precision 1 and in range (1–20)), T =

100*3 (to calculate the trust value at each time point,

predict the reliance decision and calculate the dis-

tance from the empirical data), and N = 3, then this

counts to 1 × 1 × 1 × 104×2 × 203 × 3 × 102 × 3 = 7.2 ×

1014 computation steps, which on 2.4 MHz computer

will take approx. 3.47 days. For a linear model the

computation time is about 37.5 seconds. So to vali-

date all seven models against one subject will take

10.41 days. If all subjects are validated for all seven

models in a serial fashion (one by one) on a machine

having speed 2.4 MHz then it will cost 166.66 days

to complete. Hence during the process of tuning two

approaches are followed a) to decrease granularity of

the parameters from 0.01 to 0.025 (for α, τ, β) and 1

to 2 (for σ) and secondly to use DAS-4 [1] (distrib-

uted ASCI super computer version 4) which can

distribute the validation of each of the subjects on a

separate machine in a distributed cluster. Hence 16

machines on DAS-4 have been utilized for this pur-

pose. On average these machines have provided 0.31

MHz of computation power. These steps have

speedup the process very much and the whole proc-

ess took approximately 6.19 hours on DAS-4 with

these parameters.

7.3. Validation results

From the data of 18 participants, two outliers have

been removed, which leaves a data set of 16 accura-

cies per model type (UM, LiE, LiT, LiET, LoE, LoT,

LoET and MAX).

The actual found tuned parameters per model type

per participant are too numerous to show in the paper.

Hence we only show the found accuracies.

In Fig. 8a the subjects are shown on the x-axis

while the prediction accuracies of the models are

presented on y-axis. Here it can be seen that the LiE

and LoET variants are mostly on the upper bound of

the prediction accuracy whereas the LiT, LiET, and

LoT are on the lower bound. In Fig. 8b the average

accuracy of the models over the participants is shown.

It can be seen that the LiE and LoET variant provide

better predictions while the LiT, LiET, LoE, and LoT

perform worse compared to the baseline model (UM).

In Fig. 9 the main effect of model type for accu-

racy for known data is shown. A repeated measures

analysis of variance (ANOVA) showed a significant

main effect (F(7, 105) = 61.04, p .01). A post-hoc

Bonferroni test showed that there is a significant

difference between all biased model types and the

unbiased model (UM), p 0.01, for all tests. For

models UM, LiT, LiET and LoT a significantly

higher accuracy was found for the best fit model

(MAX), p 0.01, for all tests.

Finally, for unknown data, a paired t-test showed a

significant improved accuracy of the best fit model

(M = 0.70, SD = 0.16) compared to the unbiased

model (M = 0.66, SD = 0.15), t(15) = 3.13, p 0.01.

This means that at least one of the different biased

models shows an increased capability to estimate

trust of the tested participants, also for unknown data.

8. Discussion and conclusions

In this paper, approaches have been presented that

allow for modelling biases in human trust dynamics.

In order to come to models incorporating such ap-

proaches, an existing model [11], which is often

applied (e.g., [17–19]), has been extended with addi-

tional constructs. A number of different variants have

hereby been introduced:

M. Hoogendoorn et al. / Modelling biased human trust dynamics38

AUTHOR COPY

(1) a model that strictly places the bias on the ex-

perience obtained from the trustee,

(2) a model that combines the trust and experience

and then applies the bias,

(3) a model that uses the previous trust value on

which the bias is applied.

Simulation results of the behaviour of each of the

models have been shown, as well as a comparison of

the behaviour of the models via the mutual model

mirroring method presented in [9]. Furthermore, the

resulting simulation traces have been formally ana-

lysed by means of the verification of formal proper-

ties and were shown to behave as expected. In addi-

tion, a detailed mathematical analysis has been per-

formed to investigate dynamic properties of bias-

based trust models. The properties addressed include

aspects such as when trust is increasing or decreasing,

which equilibria are possible (i.e., T(t + Δt) = T(t )),

and how the behaviour of the models is near the

equilibria, in particular whether they are attracting

and what the rate of convergence to such an equilib-

rium is. The main goal of the research presented here

is to model and validate human bias-based trust.

Therefore, an extensive validation has taken place in

which the bias-based trust models were used to de-

scribe and forecast human trust levels. In this paper,

to tailor the model to a specific human, a simple

parameter estimation technique has been used, but

more complex estimation techniques could also be

applied. The tuning technique used for the personal-

ization of trust models was inspired by the techniques

presented in [6]. The technique applied being exhaus-

tive in nature consumes a lot of computation power.

Hence, during the process of tuning two approaches

are followed a) to decrease granularity of the parame-

ters and secondly to use DAS-4 [1] which can dis-

tribute the validation of each of the subjects on sepa-

rate machine in a distributed cluster. In total 16 ma-

chines on DAS-4 have been utilized for this purpose.

These steps have speedup the process significantly:

approximately 6.19 hours on DAS-4 instead of 166

days on a personal computer.

The validation study of bias-based trust models

showed that for each participant at least one of the

different biased models has an increased capability to

estimate trust, also for unknown data. For known

data (i.e., the models were tuned to it), all of the

models are better compared to the tuned unbiased

model. The latter means that if one is able to develop

a kind of on-line tuning, the accuracies of the models

would certainly benefit. The first means that the

identification of personal characteristics might lead

to an online form of the selection of the best fit

model for unknown data, which on its turn leads to

an improved accuracy.

Within the domain of agent systems, quite some

trust models have been developed, see e.g., [13,14]

for an overview. Although the focus of this paper has

been on the design of bias-based trust models and

validation of these models, other trust models can

also be validated using the experimental data ob-

Fig. 9. Main effect of model type on accuracy.

Fig. 8. a) prediction accuracy of models across subjects, b) average

prediction accuracy of models for all subjects.

M. Hoogendoorn et al. / Modelling biased human trust dynamics 39

AUTHOR COPY

tained in combination with parameter estimation.

This is part of future work. Furthermore, other pa-

rameter adaptation methods will be explored or ex-

tended for the purpose of real-time adaptation. In

addition, we aim to implement a personal assistant

software agent that is able to monitor and balance the

functional state of the human in a timely and knowl-

edgeable manner. Also applications in different do-

mains are explorable, such as the military and air

traffic control domain.

In future, given the approach presented in this pa-

per, other models that represent human trust from the

literature, for example, addressing trust in agents as

teammates (see e.g., [3a,13,14]) could also be ex-

tended with the notion of human biases. Furthermore

it could be investigated that how far these extensions

improve the accuracy of those models.

References

[1] H. Bal, R. Bhoedjang, R. Hofman, C. Jacobs, T. Kielmann,

J. Maassen et al., The distributed ASCI supercomputer

project, SIGOPS Operating System Review 34(4) (2000),

76–96.

[2] T. Bosse, C. Jonker, L.v.d. Meij, A. Sharpanskykh, and

J. Treur, Specification and verification of dynamics in agent

models, International Journal of Cooperative Information

Systems 18 (2009), 167–193.

[3] R. Falcone and C. Castelfranchi, Trust dynamics: How trust

is influenced by direct experiences and by trust itself, in:

Proc. of the Third International Joint Conference on

Autonomous Agents and Multiagent Systems, 2004, pp. 740–

747.

[3a] X. Fan, S. Oh, M. McNeese, J. Yen, H. Cuevas, L. Strater,

and M.R. Endsley, The influence of agent reliability on trust

in human‐agent collaboration, in: Proc. of the European

Conference on Cognitive Ergonomics, 2008.

[4] L. Huff and L. Kelley, Levels of organizational trust in

individualist versus collectivist societies: A seven nation

study, Organizational Science 14 (2003), 81–90.

[5] M. Hoogendoorn, S.W. Jaffry, P.-P. van Maanen, and

J. Treur, Modeling and validation of biased human trust, in:

Proc. of the Eleventh IEEE/WIC/ACM International

Conference on Intelligent Agent Technology, O. Boissier

et al., eds, IEEE Computer Society Press, 2011, pp. 256–

263.

[6] M. Hoogendoorn, S.W. Jaffry, and J.Treur, Modeling

dynamics of relative trust of competitive information agents,

in: Proc. of the Twelfth International Workshop on

Cooperative Information Agents, M. Klusch, M. Pechoucek,

and A. Polleres, eds, Lecture Notes in Artificial Intelligence,

Vol. 5180, Springer Verlag, 2008, pp. 55–70.

[7] M. Hoogendoorn, S.W. Jaffry, and J. Treur, An adaptive

agent model estimating human trust in information sources,

in: Proc. of the Ninth IEEE/WIC/ACM International

Conference on Intelligent Agent Technology, R. Baeza-

Yates, J. Lang, S. Mitra, S. Parsons and G. Pasi, eds, IEEE

Computer Society Press, 2009, pp. 458–465.

[8] M. Hoogendoorn, S.W. Jaffry, and J. Treur, Cognitive and

neural modeling of dynamics of trust in competitive

trustees. Cognitive Systems Research, 2012, in press.

[9] S.W. Jaffry and J. Treur, Comparing a cognitive and a neural

model for relative trust dynamics, in: Proc. of Sixteenth

International Conference on Neural Information Processing,

Part I, C.S. Leung, M. Lee, and J.H. Chan, eds, Lecture

Notes in Computer Science, Vol. 5863, Springer Verlag,

2009, pp. 72–83.

[10] C.M. Jonker, J.J.P. Schalken, J. Theeuwes, and J. Treur,

Human experiments in trust dynamics, in: Proc. of the

Second International Conference on Trust Management,

Lecture Notes in Computer Science, Vol. 2995, Springer

Verlag, 2004, pp. 206–220.

[11] C.M. Jonker and J. Treur, Formal analysis of models for the

dynamics of trust based on experiences, in: Multi-Agent

System Engineering, Proc. of the Ninth European Workshop

on Modelling Autonomous Agents in a Multi-Agent World,

F.J. Garijo and M. Boman, eds, Lecture Notes in Computer

Science, Vol. 1647, Springer Verlag, 1998, pp. 221–232.

[12] P.-P.v. Maanen, T. Klos, and K.v. Dongen, Aiding human

reliance decision making using computational models of

trust, in: Proc. of the Workshop on Communication Between

Human and Artificial Agents, IEEE Computer Society Press,

2007, pp. 372–376.

[13] S. Ramchurn, D. Huynh, and N. Jennings, Trust in multi-

agent systems, The Knowledge Engineering Review 19

(2004), 1–25.

[14] J. Sabater and C. Sierra, Review on computational trust and

reputation models, Artificial Intelligence Review 24 (2005),

33–60.

[15] D.O. Sears, The person positivity bias, Journal of

Personality and Social Psychology 44 (2007), 233–250.

[16] A. Sharpanskykh and J. Treur, A temporal trace language for

formal modelling and analysis of agent systems, in:

Specification and Verification of Multi-Agent Systems,

M. Dastani, K.V. Hindriks, and J.J.Ch. Meyer, eds, Springer

Verlag, 2010, pp. 317–352.

[17] S.I. Singh and S.K. Sinha, A new trust model based on time

series prediction and Markov model, in: Proc. of the

International Conference on Information and

Communication Technologies, V.V. Das and R. Vijaykumar,

eds, Communications in Computer and Information Science,

Vol. 101, Springer Verlag, 2010, pp. 148–156.

[18] F. Skopik, D. Schall, and S. Dustdar, Modeling and mining

of dynamic trust in complex service-oriented systems,

Information Systems 35 (2010), 735–757.

[19] F.E. Walter, S. Battiston, and F. Schweitzer, Personalised

and dynamic trust in social networks, in: Proc. of the Third

ACM Conference on Recommender Systems, L. Bergman,

A. Tuzhilin, R. Burke, A. Felfernig, and L. Schmidt-Thieme,

eds, ACM Press, 2009, pp. 197–204.

[20] T. Yamagishi, N. Jin, and A.S. Miller, In-group bias and

culture of collectivism, Asian Journal of Social Psychology

1 (1998), 315–328.

M. Hoogendoorn et al. / Modelling biased human trust dynamics40