Comparison of Surveillance Strategies to Identify Undesirable Behaviour

in Multi-Agent Systems

Sarah Edenhofer

, Christopher Stifter

, Sven Tomforde

, Jan Kantert

Christian M

uller-Schloer

and J

org H

ahner

Organic Computing Group, University of Augsburg, Augsburg, Germany

Institute of Systems Engineering, Leibniz University Hannover, Hannover, Germany

Keywords:

Multi-Agent Systems, Trust, Cooperation, Coordination, Self-organising Systems, Organic Computing.

Abstract:

Open, distributed systems face the challenge to maintain an appropriate operation performance even in the

presence of bad behaving or malicious agents. A promising mechanism to counter the resulting negative

impact of such agents is to establish technical trust. In this paper, we investigate strategies to improve the

efﬁciency of trust mechanisms regarding the isolation of undesired participants by means of reputation and

accusation techniques. We demonstrate the potential beneﬁt of the developed techniques within simulations

of a Trusted Desktop Computing Grid.

1 INTRODUCTION

Open distributed systems are characterised by the

possibility to join and leave at any time. This entails

the problem that the participating elements (i.e. the

agents) are not under control of a centralised entity –

consequently, their behaviour has to be considered as

black-box. These black-boxes can contribute to the

overall system’s goal, but they can also disturb the ef-

ﬁcient operation by means of unintentionally wrong

or even intentionally malicious behaviour.

A basic mechanism to counter the resulting neg-

ative impact of such disturbances is the utilisation of

technical trust. Establishing trust relationships among

distributed agents that model the reliability and trust-

worthiness of interaction partners allows for isolat-

ing untrusted agents, see (Stegh

ofer and Reif, 2012).

In this paper, we investigate techniques to improve

the efﬁciency of isolation mechanisms on the basis

of technical trust. Therefore, we introduce and com-

pare two novel strategies for distributed surveillance:

a reputation-based and an accusation-based strategy.

Within simulations of our application scenario, we

demonstrate the potential beneﬁt of these strategies

and highlight that the duration until an effective isola-

tion of malicious elements takes place can be reduced

signiﬁcantly.

The remainder of this paper is structured as fol-

lows: In Section 2, we introduce our application sce-

nario – the Trusted Desktop Grid (TDG) – and deﬁne

agents categories and their goals. Section 3 presents

the novel strategies to isolate malicious agents us-

ing reputation and accusation techniques. Afterwards,

Section 4 evaluates the approach using simulations of

the TDG. Section 5 discusses the achieved results and

derives a statement of which technique should be used

in which cases, we present related work in Section 6.

Finally, Section 7 summarises the paper and gives an

outlook to future work.

2 TRUSTED DESKTOP GRID

We use an open, distributed Trusted Desktop

Grid (TDG) as application scenario to show and prove

the effective application of distributed algorithms as

well as Organic Computing (M

uller-Schloer et al.,

2011) methods. In this scenario, we use an open and

heterogeneous Multi-Agent System (MAS) and we do

not assume benevolence. The agents in the system

cooperate to gain an advantage. The mechanism de-

termining this cooperation is Trust. Because of the

openness of the system, different agents may try to

exploit it. They may be uncooperative, malfunction-

ing or even malicious.

An agent, which acts on behalf of the user, is sub-

mitting jobs it wants to have calculated (Klejnowski,

132

Edenhofer, S., Stifter, C., Tomforde, S., Kantert, J., Müller-Schloer, C. and Hähner, J.

Comparison of Surveillance Strategies to Identify Undesirable Behaviour in Multi-Agent Systems.

DOI: 10.5220/0005679201320140

In Proceedings of the 8th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2016) - Volume 1, pages 132-140

ISBN: 978-989-758-172-4

2014). Each job is assumed to be composed of sev-

eral independently processable work units (WU). The

agents are expected to volunteer their machines as

workers for other agents’ WU as well as to share their

resources.

2.1 Agent Goal

The beneﬁt of an agent can be measured by its

speedup σ, informally speaking its beneﬁt of having

its work processed distributively over having to pro-

cess all work on its own (in accordance with (Kle-

jnowski, 2014)). A job J is a set of WUs, which is re-

leased in time step t

rel

and completed in t

compl

, when

the last WU is ﬁnished.

The speedup σ in Equation 1 is a metric known

from multi-core systems. It is based on the assump-

tion that parallelisation helps to process a task (i.e. a

job) faster than processing it on a single core. σ is

the ratio of the time the agent would have needed to

process all WUs on its own to the real time it took to

calculate J distributedly in the system. This is why

the speedup can only be determined after the last re-

sult has been returned to the submitter.

σ =

∑

compl

sel f

−t

rel

sel f

)

∑

compl

dist

−t

rel

dist

)

(1)

In short, we can write σ

sel f

dist

with t

sel f

being the

time it would require an agent to process all WUs of

a job without cooperation, i.e. sequentially. t

dist

is the

time it takes until all WUs are computed distributedly

and the last result is returned to the submitting agent.

If no cooperation partners can be found, agents need

to calculate their own WUs. This results in a speedup

value equal to one. In general, we assume that agents

behave selﬁshly and only cooperate if they can expect

an advantage, i.e. σ > 1.

2.2 Worker and Submitter Component

Each agent is free do decide which agent it wants

to give its WUs to and for which agents it wants to

work for. Therefore, every agent has a submitter and

a worker component.

The submitter component is the scheduler of the

agent and responsible for distributing WUs. If an

agent receives a job J from the user consisting of

multiple WUs, it creates a list of suited workers, i.e.

workers it trusts. It then asks workers from this list

to cooperate and calculate WUs, until either no more

WU or no more workers are left. If all workers were

asked and still unprocessed WUs remain, the agent

calculates them on its own.

The worker component decides whether an agent

wants to work for a certain submitter. When the agent

receives a request to process a WU, it calculates its

expected reward for accepting and rejecting the WU.

If the reward of accepting the WU prevails, the agent

takes the WU, puts it in its own working queue, where

the WU remains until the agent starts to process it, i.e.

until the other WU in the queue were processed. Af-

terwards, it transfers the result back to the submitter

where the result is validated (Klejnowski, 2014). A

job is completed, if all WUs were returned to the sub-

mitter.

2.3 Global Goal

The global goal—also referred to as the system

goal—is to enable and encourage agents to cooper-

ate and thereby achieve the best possible average σ.

The systems’ focus is coordination, i.e. shaping the

environment in a way that allows for cooperation and,

thereby, leads to optimising the global goal.

2.4 Agent Types

In the context of our TDG, such disturbances are, for

example, agents that return wrong results, or agents

that refuse to work for other agents while submitting

WU to them. Behaviour like this can lead to a lower

system-speedup. In the following, we discuss differ-

ent types of agents, each behaving differently.

Adaptive Agents are cooperative. They work for

other agents which have good reputation in the sys-

tem. If, for example, the WU-queue of this agent

is saturated to capacity, the agent may reject another

WU.

Altruistic Agents accept every job, regardless of the

circumstances and cooperation-partners.

Freeriders do not work for other agents and reject

all work requests. However, they ask other agents to

work for them. This increases the overall system load

and can decrease the beneﬁt for well-behaving agents.

Egoists only pretend to work for other agents. They

accept most work requests but return fake results.

This wastes the time of other agents. If results have to

be validated, the average σ is decreased.

Sloppy Agents are cooperative but do only accept a

certain percentage of WU offered to them (Edenhofer

et al., 2015), expressed by the acceptance rate α.

2.5 Trust Metric

To overcome the problems of open, distributed MAS,

that we just described and to optimise the global goal,

we introduce a trust metric (Klejnowski et al., 2010),

Comparison of Surveillance Strategies to Identify Undesirable Behaviour in Multi-Agent Systems

133

justifying the name Trusted Desktop Grid. Agents get

ratings for their actions from their particular interac-

tion partners, representing the amount of trust earned

through this interaction.

An agent a has multiple ratings with values be-

tween −1 and 1 that it gets from other agents. The

amount of ratings k is limited to implement oblivion.

The global average of all ratings for one single agent

is called reputation. For further details see (Kantert

et al., 2015). These ratings allow to make estimations

about the future behaviour of an agent, based on its

previous actions. In our system, agents get a posi-

tive rating, if they work for other agents and a nega-

tive rating, if they reject or cancel work requests (Kle-

jnowski, 2014).

As a result, we can isolate malevolent agents and

minimise the negative impact of malicious agents (cf.

robustness in Section 2.3).

2.6 Trust Communities

To further increase the robustness and performance

of the system, we introduced an agent organisation

called Trust Community (TC). A TC consists of vari-

ous agents which greatly trust each other. The agents

are aware of their membership. A so-called Trusted

Community Manager (TCM) is elected by the agents.

It maintains the TC, preserves and optimises the com-

position and stability of the system. It is capable of

inviting members to the TC, excluding members from

the TC or assigning roles to the members for admin-

istrative reasons (for further details see (Klejnowski,

2014)). The election of the TCM can either depend

on the ID of the agent (the agent with the highest

or lowest ID is elected for TCM), or the reputation

(members with a high trust rating are more likely to

be elected as TCM).

The advantage of a TC organisation is that agents

can reduce security measures such as replication of

WU and are able to gain a better speedup. Members

of a TC can better resist attacks because they can al-

ways decide to just cooperate inside the community

and ignore the environment. Attacks within the TC

are managed by the TCM. It is capable of monitoring

a subset of the agents and, eventually, excludes them

in case of misbehaviour.

3 STRATEGIES

As described in Section 2.6, the TCM has the abil-

ity to monitor the TC members. Total surveillance is

neither desirable nor possible in an open, distributed

system with autonomous agents. Therefore, the TCM

Figure 1: The Round-Robin Strategy with step-width = 2

and S = 30 %, i.e. 3 of 9 agents are chosen (represented by

dark grey) and these agents are shifted by 2 each time step.

Figure 2: The Lottery-Based Strategy with S = 30 %. At

each timestep, 3 out of 9 agents are chosen (dark grey) as a

disjoint set to the agents chosen before (striped) until every

agent has been chosen once.

can only monitor a certain percentage of all TC mem-

bers at each time, expressed by the surveillance rate

S . It can use different surveillance strategies pre-

sented in this section to choose these S % of agents,

which we refer to as the chosen agents.

Accusation-based Strategy - An agent can be ac-

cused by another agent in case of an incident, for ex-

ample if it returns a fake result. The TCM is more

likely to observe an agent with more accusations than

an agent with less (or zero) accusations. For further

details see (Edenhofer et al., 2015).

Reputation-based Strategy - Here, an agent is more

likely to be monitored if its reputation is low in com-

parison to the other TC members’ reputation. This

concept can be realised by using, e.g. a roulette-wheel

approach that considers the available reputation val-

ues.

Random-based Strategy - The agents that are ob-

served are chosen randomly from all agents available

in each time step. Each agent has the same probability

to be selected.

Round-Robin-based Strategy - For this strategy, we

consider all TC-members as elements in a sorted list.

The chosen agents are the ﬁrst S % elements in the

ﬁrst time step. In the second time step, the chosen

agents are the next S % elements, shifted by step-

width. This is illustrated in Figure 1: the chosen

agents (dark grey) in t

are a

, a

and a

, in t

these

are a

, a

and a

, because the selection is shifted by

the step-width = 2.

Lottery-based Strategy - The agents that should be

observed are randomly chosen in each time frame but

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

134

no repetition is allowed until all agents have been ob-

served. In Figure 2, we can see that in t

agents a

, a

and a

are selected (dark grey). In t

, these agents

cannot be chosen again (striped), a

, a

and a

are

chosen. In this example, the remaining three agents

, a

and a

would be the chosen ones in t

, in t

any

agent can be selected again.

The accusation-based and reputation-based strate-

gies use additional knowledge about the agents (i.e.

their previous incidents, respectively their reputation-

value). We call these strategies quality-based. In con-

trast to that, we call the other three strategies quantity-

based, since their success highly depends on the pa-

rameter assignment of S (cf. Section 4). It should fur-

ther be noted that the knowledge acquisition equals

computational costs.

4 EVALUATION

We evaluated the performance of the surveillance

strategies using two performance metrics: the number

of exclusions (from a TC) and the average residence

time. The residence time of an agent a (t

res

) is the time

the agent spends inside the TC (see Equation 2), from

joining the TC (t

join

) to being excluded (t

exclusion

res

= t

exclusion

−t

join

(2)

During our experiments, we iterated over the parame-

ters listed in Table 1.

Table 1: Parameter for the evaluation.

parameter abbreviation

incidents before exclusion ibe

forgiveness z

acceptance (rate) α

surveillance (rate) S

ibe determines the number of incidents one agent

can commit before being excluded from the TC. In

most cases we use a value of 2 for this, because this

“three strikes” approach is a compromise between

giving autonomous agents the chance to improve their

behaviour on the one hand and having a healthy TC

on the other hand. All incidents have a timeout (in

ticks) called forgiveness z. Our experiments compare

z = 250, 1000, 5000, 10000. Lower values do not re-

sult in a signiﬁcant number of exclusions, because the

older incidents are forgiven too fast. Higher values

result in some sort of saturation and add no value to

the system – resulting in a decreased efﬁciency. The

acceptance rate α of the Sloppy Agents (SLA) ranges

from 50% to 90%, increased by 10. With lower val-

ues, the probability that an agent is invited into a TC

is too low. A value of 100% would equal the strategy

performed by Altruistic Agents (ALT).

As discussed in Section 3, we are interested in

rather low surveillance rates. Most comparisons are

using a S of 10% and 30%, though we will also have

a look at a low S of 2%, 5% and 8%. All simulations

lasted 200000 ticks with 100 agents each (30 SLA and

70 ALT). If an agent was excluded from a TC, it was

blacklisted for the remaining runtime and therefore

could not get excluded twice from the same TC. In

most upcoming sections we only allowed for one TC

during our experiments. Thereby we wanted to en-

sure a controlled environment. Nevertheless, we will

discuss some results with multiple TC in Section 4.3.

For the evaluation, we used the average of 40 runs

(respectively the average of 10 runs with multiple TC)

of one setup, as well as the standard-deviation.

In Section 4.1, we show the results for the

quantity-based strategies, the results for quality-based

strategies are introduced in Section 4.2, followed by

the results for multiple TC in Section 4.3.

4.1 Quantity-based Strategies

In this section, we compare the three quantity-based

strategies, i.e. random-based, round-robin-based, and

lottery-based, with each other. For the round-robin-

based strategy, we use a step-width of 2. Using differ-

ent step-widths does not yield a better performance.

The three strategies (cf. Figure 3) have pretty much

the same performance. Both the number of exclu-

sions and t

res

show the same trend and very sim-

ilar values. If S is increased to 30%, more agents

are excluded. Furthermore, t

res

is lower than with

S = 10% (with smaller standard deviation). The sim-

ilarity between the strategies exists with S = 10% as

well as with S = 30%. Given that these strategies per-

form that similar is due to their simplicity, we will

only use the random-based strategy for comparison

with the quality-based strategies in the following sec-

tions. Therefore, Figure 4 shows the performance of

the random-based strategy at different z levels. At

z = 250, there are nearly no exclusions and if there

are some, t

res

is pretty bad. The performance gets

better the higher z.

4.2 Quality-based Strategies

In this section, we show the evaluation of the

accusation-based and the reputation-based strategy.

Figures 6, 7 and 8 show the results for the reputation-

based strategy. Figure 6 shows the performance with

different values for z at different S levels: 10% in (a)

- (d) and 30% in (e) - (h). Similar to the random-based

Comparison of Surveillance Strategies to Identify Undesirable Behaviour in Multi-Agent Systems

135

(a) Random-based with S = 10% (b) Round-robin-based with S = 10% (c) Lottery-based with S = 10%

(d) Random-based with S = 30% (e) Round-robin-based with S = 30% (f) Lottery-based with S = 30%

Figure 3: Comparison of the quantity-based strategies at z = 10000. From left to right: results of random-, round-robin-,

lottery-based strategy. The inﬂuence of S can be seen with S = 10% (top) and S = 30% (bottom). The x-axis shows the

acceptance rate α, the left y-axis and the dashed line show t

res

, the right y-axis and the thicker solid line show the number

of exclusions.

(a) Random-based, z = 250, S = 30. (b) Random-based, z = 1000, S = 30. (c) Random-based, z = 5000, S = 30.

Figure 4: Random-based strategy with S = 30, ibe = 2 and z = 250, 1000, 5000. The x-axis shows the acceptance rate α, the

left y-axis and the dashed line show t

res

, the right y-axis and the thicker solid line show the number of exclusions.

strategy the performance increases with increasing z.

Be aware that t

res

is not meaningful, if the number

of exclusions is almost zero (cf. for example with Fig-

ure 6 (a)). If z is further increased to 10000, the out-

come stagnates. Changing S from 10% to 30% does

not yield a huge improvement in performance. If we

lower S (cf. Figure 7), we eventually see a drop in

performance at S = 2. An increase of the ibe value

creates a lower outcome (cf. Figure 8).

For the accusation-based strategy we show the re-

sults for different values for z at S = 10. The strat-

egy shows very similar values to the reputation-based

strategy, while showing the same trends if varying the

parameters.

4.3 Evaluation With Multiple TC

During the above mentioned evaluations we only al-

lowed for one TC. In Figure 9, we show the results

of the accusation- and reputation-based strategy, if

multiple TC are allowed. The trends – with differ-

ent parameter assignments – already described remain

valid. If we compare the values of the accusation-

based strategy to those of the reputation-based one,

we can see that the results of the reputation-based

strategy are signiﬁcantly better.

5 DISCUSSION

As mentioned in Section 4, our experiments iterated

over the parameters listed in Table 1. As we can see

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

136

(a) Accusation-based with z = 250 (b) Accusation-based with z = 1000 (c) Accusation-based with z = 5000

Figure 5: Comparison of the accusation-based strategy’s performance with different values for z (left to right) at S = 10. The

x-axis shows the acceptance rate α, the left y-axis and the dashed line show t

res

, the right y-axis and the thicker solid line

show the number of exclusions.

(a) Rep.-based with z = 250, S = 10% (b) Rep.-based with z = 1000, S = 10% (c) Rep.-based with z = 5000, S = 10%

(d) Rep.-based with z = 250, S = 30% (e) Rep.-based with z = 1000, S = 30% (f) Rep.-based with z = 5000, S = 30%

Figure 6: Comparison of the reputation-based strategy’s performance with different values for z (left to right) at different S

levels: (a) - (d) 10%, (e) - (h) 30%. The x-axis shows the acceptance rate α, the left y-axis and the dashed line show t

res

the right y-axis and the thicker solid line show the number of exclusions.

(a) Reputation-based with S = 2% (b) Reputation-based with S = 5% (c) Reputation-based with S = 8%

Figure 7: Reputation-based strategy with z = 10000 at different S levels: 2%, 5%, 8% (left to right). The x-axis shows the

acceptance rate α, the left y-axis and the dashed line show t

res

, the right y-axis and the thicker solid line show the number

of exclusions.

in all ﬁgures, the performance of the strategies de-

pends on the value for α. An increasing α has two

major effects: First, the number of SLA in the TC

gets higher. This simply leads to more exclusions.

Second, the higher α, the less incidents. This means,

the time between two incidents of the same agent in-

creases. Eventually, this time is too long in relation

to the current z. Therefore, the older incidents are

Comparison of Surveillance Strategies to Identify Undesirable Behaviour in Multi-Agent Systems

137

(a) Rep.-based with ibe = 0 (b) Rep.-based with ibe = 1 (c) Rep.-based with ibe = 2 (d) Rep.-based with ibe = 3

Figure 8: Comparison of the reputation-based strategy’s performance with z = 5000, S = 10 and different ibe values: 0, 1, 2, 3

(left to right). The x-axis shows the acceptance rate α, the left y-axis and the dashed line show t

res

, the right y-axis and the

thicker solid line show the number of exclusions.

(a) Accusation-based, z = 250 (b) Accusation-based, z = 1000 (c) Accusation-based, z = 5000

(d) Reputation-based, z = 250 (e) Reputation-based z = 1000 (f) Reputation-based z = 5000

Figure 9: Comparison of the accusation-based ((a) - (c)) and the reputation-based strategy ((d) - (f)) with different values for

z (left to right: 250, 1000, 5000) at S = 10, if multiple TC are allowed. The x-axis shows the acceptance rate α, the left y-axis

and the dashed line show t

res

, the right y-axis and the thicker solid line show the number of exclusions.

already forgiven when the agent commits another vi-

olation. This second effect seems to dominate when-

ever the performance drops although α is increasing

(for example at around 70% in Figure 4 (c)).

In Figures 4, 5, 6, and 9, we compare the perfor-

mance for varying z. The results have in common

that at z = 250 the performance is pretty bad, but

progressively getting better with increasing z. This,

again, is due to z being too small: when the last in-

cident, which is necessary to exclude an agent, is ob-

served, the ﬁrst incident is already forgiven. For the

same reasons the performance gets worse with higher

ibe (cf. Figure 8): the higher the value of ibe, the

longer z must be. In case of increasing z, the out-

come stagnates at around z ≈ 5000. A higher z has

no additional beneﬁt, because all agents are already

excluded at lower z values.

Additionally, we showed the impact of different

S levels in Figures 3, 6, and especially 7. It is plau-

sible that the performance gets better with higher S ,

particularly resulting in a higher number of exclu-

sions. If S is too low, it takes too long to repeat-

edly observe every SLA and consequently, again, z

is too low. The outcome stagnates with increasing S ,

because there are no more agents to exclude. This

stagnation arises at different S for the quantity-based

and quality-based strategies. This is due to the way

these groups of strategies choose the agents they ob-

serve: the quantity-based strategies choose indepen-

dently from the agent’s identity. This means, the S %

agents are chosen from the total number of agents of

the TC. In contrast, if using a quality-based strategy,

the S % agents are primarily chosen from the SLA

subset of agents.

Comparing the accusation- with the reputation-

based strategy is especially interesting if we allow for

multiple TC (cf. Section 4.3 and Figure 9). While

the results are pretty similar if only one TC is al-

lowed (cf. Figures 5 and 6 (a) - (c)), we can see a

signiﬁcant advantage of the reputation-based strategy

over the accusation-based one in regards to both the

number of exclusions and t

res

. This is particularly

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

138

true for higher α. If multiple TC are allowed, an ex-

cluded agent cannot join the same TC again, but is

allowed to join any other (this is also the reason why

there are more exclusions, considering the changed

value range in Figure 9). To explain why reputation-

based performs better than accusation-based, we have

to consider the peculiarities of these two strategies.

Both use a form of history when judging which agents

should be observed: the accusation-based strategy

is more likely to observe agents which accumulated

more accusations in the past than agents that did not

do so. The reputation-based strategy is more likely to

observe agents with a low reputation (relative to the

other TC members). If an agent is excluded from a

TC, its accusations are reset: the TCM of the new TC,

where this agent becomes a member, does not know

about former accusations in other TC. In contrast to

that, when the TCM uses the reputation-based strat-

egy, there is already some information available for

the TCM in form of the agent’s reputation. Therefore,

from the moment the agent enters the new TC, it is

already more likely to be observed. Thus, the results

focussing on the number of exclusions as well as t

res

of the reputation-based strategy are superior to those

of the accusation-based strategy.

6 RELATED WORK

Our application scenario described in Section 2 is a

Desktop Grid System. We model our grid nodes as

agents, which can be seen as black boxes. Thereby,

we cannot observe the internal state. Thus, their ac-

tions and behaviour can only be predicted with un-

certainty (Hewitt, 1991). Our TDG supports Bag-of-

Tasks application (Anglano et al., 2006). A classiﬁ-

cation and taxonomy of Desktop Grid Systems can be

found in (Choi et al., 2007), respectively (Choi et al.,

2008).

Desktop Grid Systems are used to share resources

between multiple administrative authorities. One ex-

ample for a peer-to-peer based system is the Share-

Grid Project (Anglano et al., 2008). A second ap-

proach is the Organic Grid, a peer-to-peer based

approach with decentralised scheduling (Chakravarti

et al., 2004). Compared to our system, these ap-

proaches assume benevolence (Wang and Vassileva,

2004), i.e. that there are no malicious agents partic-

ipating and misbehaving. Another approach is the

open source Berkeley Open Infrastructure for Net-

work Computing Project (BOINC) (Anderson and

Fedak, 2006) or XtremWeb (Fedak et al., 2001),

which aims at setting up a Global Computing ap-

plication and “harvest[ing] the idle time of Inter-

net connected computers which may be widely dis-

tributed across the world, to run a very large and

distributed application” with an ad-hoc veriﬁcation

process for participating computers. We introduce a

trust metric (see Section 2.5) with a reputation sys-

tem to cope with the problem of misbehaving agents.

A panoramic view on computational trust in Multi

Agent Systems can be found in (Ramchurn et al.,

2004), (Castelfranchi and Falcone, 2010), or (Sabater

and Sierra, 2005). Sabotage-tolerance and distributed

trust management in Desktop Grid Systems was eval-

uated in (Domingues et al., 2007). Here, mechanisms

for sabotage detection are presented, but proposed for

a paradigm of volunteer-based computing. Trust is

also used in other disciplines such as philosophy (Kar-

lins and Abelson, 1959), psychology (Hume, 1739),

or sociology (Buskens, 1998).

7 CONCLUSION

In this paper, we presented several strategies to ﬁnd

and exclude malicious or bad behaving agents from

open, distributed multi-agent systems. We have to

differentiate between quantity-based strategies (such

as the random-based, the round-robin-based, and the

Lottery-based strategy) and quality-based strategies

(like the accusation-based and the reputation-based

strategy). To show the advantages and disadvantages

of these strategies, we implemented them in our appli-

cation scenario, the Trusted Desktop Grid (TDG). In

several evaluations we showed the inﬂuence of the pa-

rameters acceptance rate (how many work units does

one agent accept?), surveillance rate (how much bud-

get do we have for surveillance, how many agents

can we observe within a certain period of time?), the

forgiveness (when will a former incident be forgot-

ten?), as well as the incidents before exclusion. We

found out that in most cases (especially at low surveil-

lance levels) the quality-based strategies perform bet-

ter than the quantity-based strategies. For future work

we plan to further improve the TDG and the surveil-

lance in multi-agent systems, e.g. by establishing a

dynamic surveillance. This means, the choice of the

current strategy used for surveillance will depend on

a dynamically changing surveillance budget to make

the system more realistic and life-like.

REFERENCES

Anderson, D. and Fedak, G. (2006). The Computational and

Storage Potential of Volunteer Computing. In Proc. of

CCGRID 2006, pages 73–80. IEEE.

Comparison of Surveillance Strategies to Identify Undesirable Behaviour in Multi-Agent Systems

139

Anglano, C., Brevik, J., Canonico, M., Nurmi, D., and Wol-

ski, R. (2006). Fault-aware Scheduling for Bag-of-

Tasks Applications on Desktop Grids. In Proc. of

GRID 2006, pages 56–63. IEEE.

Anglano, C., Canonico, M., Guazzone, M., Botta, M., Ra-

bellino, S., Arena, S., and Girardi, G. (2008). Peer-to-

Peer Desktop Grids in the Real World: The ShareGrid

Project. Proc. of CCGrid 2008, pages 609–614.

Buskens, V. (1998). The Social Structure of Trust. Social

Networks, 20(3):265 – 289.

Castelfranchi, C. and Falcone, R. (2010). Trust Theory:

A Socio-Cognitive and Computational Model, vol-

ume 18. John Wiley & Sons.

Chakravarti, A. J., Baumgartner, G., and Lauria, M. (2004).

Application-Speciﬁc Scheduling for the Organic Grid.

In Proc. of GRID 2004 Workshops, pages 146–155,

Washington, DC, USA. IEEE.

Choi, S., Buyya, R., Kim, H., and Byun, E. (2008). A

Taxonomy of Desktop Grids and its Mapping to State

of the Art Systems. Technical report, Grid Comput-

ing and Dist. Sys. Laboratory, The University of Mel-

bourne.

Choi, S., Kim, H., Byun, E., Baik, M., Kim, S., Park, C.,

and Hwang, C. (2007). Characterizing and Classifying

Desktop Grid. In Proc. of CCGRID 2007, pages 743–

748.

Domingues, P., Sousa, B., and Moura Silva, L. (2007).

Sabotage-Tolerance and Trustmanagement in Desktop

Grid Computing. In Future Gener. Comput. Syst. 23,

7, pages 904–912.

Edenhofer, S., Stifter, C., J

anen, U., Kantert, J., Tomforde,

S., H

ahner, J., and M

uller-Schloer, C. (2015). An

Accusation-Based Strategy to Handle Undesirable Be-

haviour in Multi-Agent Systems. In 2015 IEEE Eighth

International Conference on Autonomic Computing

Workshops (ICACW).

Fedak, G., Germain, C., Neri, V., and Cappello, F. (2001).

Xtremweb: A Generic Global Computing System. In

Cluster Computing and the Grid, 2001. Proceedings.

First IEEE/ACM International Symposium on, pages

582–587.

Hewitt, C. (1991). Open Information Systems Semantics

for Distributed Artiﬁcial Intelligence. Artiﬁcial intel-

ligence, 47(1):79–106.

Hume, D. (1739). A Treatise of Human Nature. (Oxford

Clarendon Press.

Kantert, J., Edenhofer, S., Tomforde, S., and M

uller-

Schloer, C. (2015). Representation of Trust and Repu-

tation in Self-Managed Computing Systems. In IEEE

13th International Conference on Dependable, Au-

tonomic and Secure Computing, DASC 2015, pages

1827–1834, Liverpool, UK. IEEE.

Karlins, M. and Abelson, H. (1959). Persuasion: How

Opinions and Attitudes are Changed. Springer Pub.

Co.

Klejnowski, L. (2014). Trusted Community : A Novel Mul-

tiagent Organisation for Open Distributed Systems.

PhD thesis, Leibniz Universit

at Hannover.

Klejnowski, L., Bernard, Y., H

ahner, J., and M

uller-

Schloer, C. (2010). An Architecture for Trust-

Adaptive Agents. In Proc. of SASO Workshops, pages

178–183.

uller-Schloer, C., Schmeck, H., and Ungerer, T. (2011).

Organic Computing - A Paradigm Shift for Complex

Systems. Springer.

Ramchurn, S. D., Huynh, D., and Jennings, N. R. (2004).

Trust in Multi-agent Systems. Knowl. Eng. Rev.,

19(1):1–25.

Sabater, J. and Sierra, C. (2005). Review on computational

trust and reputation models. Artiﬁcial Intelligence Re-

view, 24(1):33–60.

Stegh

ofer, J.-P. and Reif, W. (2012). Die Guten, die B

osen

und die Vertrauensw

urdigen Vertrauen im Organic

Computing. Informatik-Spektrum, 35(2):119–131.

Wang, Y. and Vassileva, J. (2004). Trust-Based Community

Formation in Peer-to-Peer File Sharing Networks. In

Proc. on Web Intelligence, pages 341–348.

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

140