AGENT COALITION FORMATION VIA INDUCING TRUST

RATIO

Osama Ismail

Central Laboratory for Agricultural Expert Systems, Giza, Egypt

Samhaa R. El-Beltagy, Reem Bahgat

Faculty of Computers and Information, Cairo University, Giza, Egypt

Ahmed Rafea

Department of Computer Science, The American University in Cairo, Cairo, Egypt

Keywords: Agents, Genetic algorithms, Trust models.

Abstract: This work presents a model for assigning trust values to agents operating within a collaborative multi-agent

system. The model enables agents to assess the trustworthiness of their peers, and thus, to be able to select

reliable ones for cooperation and coalition formation. In this work, the performance of a group of agents – a

team – that collaborate to achieve a shared goal where the individual contribution of each agent is unknown,

is evaluated. The work thus aims to present a reliable method for calculating a trust value for agents

involved in teamwork. More specifically, this research presents a model – called Inducing the Trust Ratio

Model - for evaluating the individual trustworthiness of a group of agents. Toward this end, the model

makes use of genetic algorithms to induce the trust ratio of each coalition member. Empirical analysis is

undertaken to evaluate the effectiveness of this model.

1 INTRODUCTION

The work presented herein aims to augment a multi-

agent system with a trust model which enables the

collaborating agents to select peers that have the best

performance for future collaboration purposes. The

testbed used for experimenting with the developed

model is the Collaborative Expert Agents System

(CEAS) Architecture (Ismael et al., 07) in which

cooperation between heterogeneous knowledge

based systems can be achieved. Focus is placed on

developing a trust model capable of quantifying a

trust measure for individual agents working in teams

within multi-agent systems. Obviously, it is in the

best interest of the truster to delegate a task in a way

that maximizes the probability of the task being

completed with the highest possible quality of

service. Thus, agents must attempt to minimize the

risk of failure by choosing trustworthy resources. To

do so, agents must be able to accurately assess and

compare the trustworthiness (i.e. the expected

performance) of potential Provider Service Agents.

Previous work has introduced a variety of trust

models based on different criteria to derive the

trustworthiness of a single agent (e.g. Falcone et al.,

01; 03; Jensen et al., 04; Ramchurn et al., 04; Dong,

06). However, in these models the trust value is

calculated based on the individual performance of

each agent within a multi-agent system. There are

cases when the individual performance of an agent

can not be determined in a straight forward manner.

This is the case for example when an agent is

collaborating with other agents for achieving a

certain task, and where the result of the collaboration

can be evaluated but independent evaluation of the

output of each agent is not possible, For example,

suppose there is a particular problem (task) that

requires the collaboration of different agents to be

accomplished. A team of agents formulates solutions

by each tackling (one or more) sub-problems and

synthesizing these sub-problem solutions into an

469

Ismail O., El-Beltagy S., Bahgat R. and Rafea A. (2009).

AGENT COALITION FORMATION VIA INDUCING TRUST RATIO.

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence, pages 469-474

DOI: 10.5220/0001657504690474

 SciTePress

overall solution. Therefore, the final result is the

outcome of the collaboration. This result is

evaluated according to its quality and represents an

assessment of the performance of the team as a

whole. So, evaluation of this result is an evaluation

of the teamwork rather than of independent agents

that contributed to the formulation of this result.

Assuming that the output can be assigned a trust

value, the problem addressed by this work is how to

distribute this value between participating agents

according to the contribution of each. Since each

agent member in the team has particular capabilities,

each member may participate by a different ratio in

executing the allocated task. Therefore, we need to

provide a mechanism by which individual agents

within a team can be assigned a trust ratio according

to each individual agent’s capabilities; this ratio will

represent the real trustworthiness of each agent as

closely as possible.

Towards this end we investigate the use of

genetic algorithms for inducing a trust ratio for each

agent coalition member.

This paper is organized as follow: section 2

presents the proposed model – Inducing the Trust

Ratio Model (ITRM). Section 3 presents the

empirical evaluation of the proposed model while

section 4 concludes this paper and outlines some

future research directions.

2 INDUCING THE TRUST RATIO

MODEL

In this work we concentrate on efficient task

allocation through coalition formation as a means for

cooperating agents. So, the issue of which agents to

trust when forming a coalition becomes quite

important. We propose a solution to this problem

through the use of what we’ve called Inducing the

Trust Ratio Model (ITRM). To build and utilize the

model, three phases are undergone: the exploration

phase, the inducing phase and the refinement phase.

In the exploration phase, a set of test cases is

randomly generated and presented to possible agent

teams. The performance of each team is then

evaluated. In the inducing phase, exploration phase

results are analyzed and processed through the use

of a genetic algorithm. The output of this phase is a

trust value for each agent. The refinement phase is a

phase that is always active after completion of the

first two phases. In this phase agents classified as

low trust agents and agents that join the multi-agent

system are periodically re-evaluated so as to allow

such agents the chance to improve their performance

and for their trust values to be adjusted accordingly.

2.1 The Exploration Phase

In this phase, the primary purpose is to explore the

performance of a community of agents through

presenting the multi-agent system with a real task

and allowing it to form different teams to run on

randomly generated real cases for the purpose of

completing that task. The results are then assessed

through system users. So, for each case the system

will select the collaborating agents (team) from a

number of potential agents that provide the same

required service but with different qualities - in a

random way - allowing the system to learn about the

performance of an unknown provider (i.e. exploring

the provider population). For each of these cases a

system user will provide a rating as to the quality of

the teamwork. The system user in this sense acts as

an arbitrator judging the outcome against what s/he

knows should be the result. This for example can

take the form of different predefined criteria. This

rating is then recorded together with the name of the

team of participating agents in a transactions

database, for future trust evaluation in the next

phase. For simplicity we denote the name of the

agents with the name of the service they provide and

a suffix number (so agent x1 is a provider of service

x while agent z3 is a provider for service z).

Assuming we have a task that requires the

invocation of three services (hence the collaboration

of three agents), various agent combinations can be

formed for achieving this task. A system user

evaluates the output of these teams by giving it a

value which represents its quality or the trust value

for the teamwork. So, the rating for the teamwork’s

performance denotes the trust value for the entire

team and represents the sum of the participating

agents’ performances. Therefore, the collective trust

of a team can be represented by equation 1.

ct =

∈

. tr

(1)

where, ct represents the collective trust value, w

represent the weight given to agent x (determined

according to each agent’s role), tr

represents the

trust ratio that belongs to agent x, t is the set of

participating agents in the team and n is the number

of elements in the set t. For simplicity, we assume

that all agents in the domain have the same weight.

This equation is used in the next phase.

ICAART 2009 - International Conference on Agents and Artificial Intelligence

470

2.2 The Inducing Phase

The goal of this phase is to assign a trust ratio to

each agent based on a examples obtained from the

exploration phase. Since, coalition members may

differ and the space of the case's possibilities is vast,

exploring each possible case as outlined in the

previous phase with each possible team, would be

impossible. Here, we need a suitable mathematical

model that considers the set of tuples in the

transactions database as a set of simulations

equations as represented by equation 1, and resolves

the different trust values for each agent by solving

these equations simultaneously. To model this

problem, a genetic algorithm was employed. An

implementation of a genetic algorithm begins with a

population of (typically random) chromosomes. In

this case, a chromosome is a collection of trust

values; and each gene represents an agent’s

trustworthiness. Each chromosome in the population

is evaluated and ranked according to its relative

strength within the population by applying its values

on all equations. As stated before, each tuple can be

represented as an equation stored in the transactions

database. The goal from the evaluation of the

chromosome is to calculate its fitness by counting

the number of equations satisfied under this

chromosome’s values. Thus, a random population of

potential solutions is created, then each one is tested

for success, selecting the best chromosome to pass

on their 'genes' to the next generation, including

slight mutations to introduce variation. The process

is repeated until the program evolves a workable

solution. After reaching a predefined threshold that

represents the minimum accepted fitness, a

chromosome is selected to represent the optimal

solution to the problem being solved. The values of

the best chromosome represent the trust ratio of each

agent in the domain. This trust ratio is stored in the

Service Agent Profiling

database, which can be used

to enable the system to select agents that have the

best performance for future collaboration.

In the selected testbed, the Server Agent (as

illustrated in the CEAS architecture in will be the

responsible unit that manages the Service Agent

Profiling database through maintaining a record of

trustworthiness for each agent in the platform. The

Service Agent Profiling database forms the primary

source for selecting partners, and is itself updated

periodically as will be described in the next phase.

2.3 The Refinement Phase

We cannot assume that an agent’s behavior will

remain constant over time since its performance may

alter (for better or worse) over time. If a truster

knows that certain interacting partners provide an

acceptable level of service, they might never choose

to interact with any other agent that they know less

about. This attitude may mean that new agents never

get a foothold in the environment, even if they offer

a better service than other established agents. The

goal of this phase is thus to inform low performing

agents of their trust rating within the system so as to

allow them to improve their performance, and to

periodically re-evaluate the performance of agents

within the system so as to update any changes and

assign trust values to new agents. Towards the

fulfillment of the first part of this goal, the Server

Agent navigates the Service Agent Profiling

database, selects agents who have the lowest

performance (i.e. those whose trust value falls below

an acceptable level of performance) in each domain

of service, and sends them an inform message telling

them that they have the lowest trustworthiness.

Towards the fulfillment of the later, the system

periodically monitors the environment, taking into

account environment changes, such as new coming

agents, withdrawal or disappearance of previously

existing working agents, agents with improved

performance or agents with better performance than

the current set of peers previously tested. In other

words, in this phase the cycle of inducing trust ratios

is restarted from the beginning with the Service

Agent Profiling database being updated based on the

new outcomes.

3 EVALUATION

This section presents the results of an experiment

conducted in a collaborative knowledge-agents

environment with the aim of evaluating the proposed

approach. Empirical evaluation is used as the

method of measurement because it allows us to

assess the performance of a trust model in terms of

how much benefit it can bring to its users. This

requires us to compare ITRM's performance with

that of another model. But, inducing the trust ratio

for each coalition member –to our knowledge - has

not been explicitly addressed within the field of

multi-agent systems yet. Therefore, we could not

apply this type of comparison. Instead, we compare

the performance of the CEAS equipped with the

AGENT COALITION FORMATION VIA INDUCING TRUST RATIO

471

ITRM with the performance of the CEAS with no

trust model.

3.1 Description of the Testbed

In the selected testbed, there are a number of

competing Provider Service Agents that can fulfill a

particular task, with each providing a different

quality of service. Without loss of generality, and in

order to reduce the complexity of the testbed's

environment, it is assumed that there are four

distinct set of services in the testbed, called:

irrigation services, soil services, water services, and

climate services, which are provided by the

following Expert Agents respectively: Irrigation

Expert Agents, Soil Expert Agents, Water Expert

Agents, and Climate Expert Agents. Hence, there is

more than one Expert Agent which offers the same

service, but which vary in performance.

The objective here is to conduct an experiment

that demonstrates the capabilities of the system in

building an irrigation scheduling table. We take a

practical and experimental approach in our

investigation. To this end we adopt a concrete

example in the agriculture domain within which to

test and evaluate the ITRM. However the ITRM is

not restricted to this or any particular application

area.

3.2 Evaluation Tools

Two different tools have been developed to facilitate

and expedite the evaluation process, the first is the

TCGM and the second is the ESS. A Test Cases

Generator Module (TCGM) was implemented for

automatically generating a set of random test cases

for different knowledge base components according

to the contents of the knowledge base components

under consideration (Rule Cluster File). These test

cases are then stored in a composite Test Cases

Library to be used in the execution stage. Figure 1,

illustrates the overall structure of the proposed

model. The main goal for implementing this module

is to speed up the recommendation process

(Exploration Phase – as illustrated previously in

Section 2.1) through the use of a preprocessing

module which is based on three steps: random test

case generation, random test case execution and

automatic test case evaluation. It is worth noting that

the behavior obtained during this phase is

representative of the behavior to be expected during

actual interaction.

Figure 1: Overall structure of the TCGM.

To evaluate the results of the system, we’ve

developed an Expert Simulator System (ESS) whose

role is to evaluate system results. The mechanism

through which the ESS works is: The ESS is

equipped with a Knowledge Based System (KBS),

which produces what could be considered perfect

solutions to cases under consideration; these

represent the base-line for comparison. Therefore, to

evaluate the result of a specific case, the system

sends the case parameter(s) and the result to the

ESS. The ESS processes the case based on its

received parameter(s). The result of this evaluation

is a number denoting how closely the obtained result

is to the baseline.

3.3 The Experiment

The agricultural domain is one in which numerous

successful expert systems have been developed (e.g

Rafea and Mahmoud, 2001). We were able to obtain

twelve pre-existing expert systems for irrigation,

soil, water, climate systems in agriculture and

transform them into a community of cooperating

agents. These agents are divided as follows: 3

Irrigation Expert Agents, 3 Soil Expert Agents, three

Water Expert Agents, and 3 Climate Expert Agents.

For the purpose of comparison, we implemented

a version of the testbed without the facilities offered

by the ITRM and then another with these facilities.

To differentiate between the system built with a trust

model and the one without, we will refer to the

former as CEAS+ and the latter as CEAS-. Our

empirical evaluation consists of a series of

simulations tailored to show the developed model’s

ICAART 2009 - International Conference on Agents and Artificial Intelligence

472

performance. In each experiment there are two

stages: in the first stage the system generates a set of

random test cases (training cases), and presents the

multi-agent system (CEAS-) with the task of solving

these cases. In this stage, the system selects an agent

team randomly and without use of the trust model.

The collective performance for each team in the

platform is then automatically calculated (by using

the ESS). In the second stage the same set of test

cases is presented to CEAS+ where team formation

is based on previously recorded trust values and the

collective performance for the selected team is

calculated also using the ESS. The results for each

case are then compared.

The performance of each group of agents in

terms of utility gain is plotted on a chart to show the

trend of performance change. The x axis represents

the number of interactions (number of cases) and the

y axis represents the effectiveness percentage

(performance). When the number of interactions is

set to forty, the effectiveness average for CEAS+ is

about 93.03 %, indicating that the system can

achieve a stable and high performance at this point

while the effectiveness average for CEAS- is about

86.19 % for the same number of cases as illustrated

in Figure 2.

Figure 2: Effectiveness performance based on 40 previous

test cases.

When the number of interactions is set to sixty,

the effectiveness average for CEAS+ is about 93.02

%, while the effectiveness average for CEAS- is

about 86.26 %, as illustrated in Figure 3. When the

number of interactions is set to eighty, the

effectiveness average for CEAS+ is about 93.09 %,

while the effectiveness average for CEAS- is about

84.85 %, as illustrated in Figure 4. When the number

of interactions is set to one hundred, the

effectiveness average for CEAS+ is about 92.99 %,

while the effectiveness average for CEAS- is about

85.78 %, as illustrated in Figure 5.

Figure 3: Effectiveness performance based on 60 previous

test cases.

Figure 4: Effectiveness performance based on 80 previous

test cases.

Figure 5: Effectiveness performance based on 100

previous test cases.

3.4 Hypothesis Testing

A mere comparison of the performance of the two

systems does not allow us to conclude that one

system performs better than the others in all cases.

This is because the population of possible situations

is infinitely large and the results from one

experiment are only from a small sample of that

population. Given this problem, statistical inference

techniques should be used since they allow us to

draw a conclusion about an unseen population given

a relatively small sample. To the extent that a

sample is representative of the population from

which it is drawn, statistical inference permits

generalizations of conclusions beyond the sample

(Cohen, 95). The hypothesis testing method as a

statistical inference technique, allows us to confirm

Effectiveness Performance

100

1 3 5 7 9 111315171921232527293133353739

Interactions

Performance

CEAS+ CEAS-

Effectiveness Performance

100

1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58

Interactions

Performance

CEAS+ CEAS-

Effectiveness Performance

100

1 5 9 1317212529333741454953576165697377

Interactions

Performance

CEAS+ CEAS-

Effectiveness Performance

100

1 7 13 19 25 31 37 43 49 55 61 67 73 79 85 91 97

Interactions

Performance

CEAS+ CEAS-

AGENT COALITION FORMATION VIA INDUCING TRUST RATIO

473

with a predefined confidence level, whether the

difference of the two means of the two sample

groups' performance, actually indicates that one

system has higher performance than the other, and

hence, eliminate the random factor in selecting the

samples.

Table 1: Terms used in the hypothesis testing procedure.

Term Definition

CEAS+

CEAS-

CEAS+

CEAS-

The number of interactions chosen as the test

period (number of test cases).

The mean performance of a sample of agents

using ITRM after their n

interaction.

The mean performance of a sample of agents

using no trust model after their n

interaction.

The standard deviation of the performance

sample of CEAS+.

The standard deviation of the performance

sample of CEAS-.

The result of carrying out the hypothesis testing

procedure for different test periods (i.e. 10, 20, 30,

40, 60, 80, and 100 interaction) is illustrated in

Table 2.

Table 2: Hypothesis testing results.

Number of

CEAS+

CEAS-

CEAS+

CEAS-

P-value

10 87.77 84.67 1.47 4.45 1.479 4.96 2.1 0.045

20 91.13 84.71 0.84 4.91 1.114 16.32 5.8 1.46E-5

30 92.92 82.74 0.65 5.83 1.071 26.07 9.5 3.02E-10

40 93.03 86.19 0.63 5.20 0.829 57.25 8.3 1.32E-11

60 93.02 86.26 0.71 4.55 0.595 297.5 11.4 2.16E-25

80 93.09 84.85 0.98 5.11 0.585 585 14.1 2.59E-39

100 92.99 85.78 0.68 5.08 0.513 732.86 14.1 3.32E-40

Since, the P-value for all cases is less than the

significance level (0.05), we cannot accept the null

hypothesis. Therefore, this table shows that the

corresponding hypothesis tests conclude that the

CEAS+ outperforms the CEAS- and that the

performance difference is statistically significant

(using the confidence level of 95%).

4 CONCLUSIONS

Previous work addressing trust, has investigated

active trust, but passive trust has not been explicitly

addressed within the field of multi-agent systems to

date. In active trust, the performance of individual

agents from various perspectives is evaluated using

various sources of trust information, such as, direct

interaction or through witness reports. But, in such

cases, the agent that is meant to be evaluated is

known in advance. In passive trust (addressed by

this work), the performance of an agent within a

group of agents that collaborate to achieve a shared

goal is what is being evaluated. To do so, a trust

ratio for each agent in the team is induced. The

presented model for achieving this task: Inducing the

Trust Ratio Model (ITRM) is thus a novel model for

trust evaluation that is specifically designed for

general application in multi-agent systems. In order

to verify the claim that this model is both effective

and useful, empirical evaluation was carried out.

Through this evaluation it was demonstrated that

agents using the trust model - ITRM - provided by

CEAS are able to select reliable partners for

interactions and, thus, obtain better utility gain

compared to those using no trust measure.

REFERENCES

Cohen, P., 1995. Empirical Methods for Artificial

Intelligence, The MIT Press.

Dong, H., Jennings, N., Shadbolt, N., 2006. An Integrated

Trust and Reputation Model for Open Multi-Agent

Systems. In Journal of Autonomous Agents and Multi-

Agent Systems.

Falcone, R., Barber, S., Korba, L., Singh, M., (editors)

2003. Trust, Reputation and Security: Theories and

Practice, Volume 2631 of Lecture Notes in Computer

Science. Springer-Verlag Berlin Heidelberg.

Falcone, R., Singh, M., Tan, Y., (editors) 2001. Trust in

Cyber-Societies: Integrating the Human and Artificial

Perspectives, Volume 2246 of Lecture Notes in

Computer Science. Springer-Verlag Berlin Heidelberg.

Ismael, O., El-Beltagy, S., Bahgat, R., Rafea, A., 2007.

Collaborative Knowledge Based System through

Multi-Agent Technology. In the 42

Annual

Conference on Statistics, Computer Science, and

Operations Research. Institute of Statistical Studies

and Research, Cairo University.

Jensen, C., Poslad, S., Dimitrakos, T., (editors) 2004.

Trust Management, Volume 2995 of Lecture Notes in

Computer Science. Springer-Verlag Berlin Heidelberg.

Proceedings of the Second International Conference,

iTrust 2004 Oxford, UK, March 29 – April 1, 2004.

Rafea, A., Mahmoud, M., 2001. The Evaluation and

Impact of NEPER Wheat Expert System. Fourth

International Workshop on Artificial Intelligence in

Agriculture IFAC/CIGR, Budapest, Hungary.

Ramchurn, D., Dong, T., Jennings, R., 2004. Trust in

Multi-Agent Systems. The knowledge engineering

review, 19(1): 1-25, March 2004.

ICAART 2009 - International Conference on Agents and Artificial Intelligence

474