AN ADAPTIVE AND DEPENDABLE SYSTEM CONSIDERING

INTERDEPENDENCE BETWEEN AGENTS

Keinosuke Matsumoto, Tomoaki Maruo, Akifumi Tanimoto and Naoki Mori

Graduate School of Engineering, Osaka Prefecture University, 1-1 Gakuen-cho, Nakaku, Sakai, Osaka, Japan

Keywords: Dependability, Interdependence graph, Monitoring, Replication, Intelligent systems.

Abstract: A multiagent system (MAS) has recently gained public attention as a method to solve competition and

cooperation in distributed systems. However, MAS’s vulnerability due to the propagation of failure prevents

it from being applied to a large-scale system. This paper proposes an approach to improve the efficiency and

reliability of distributed systems. The approach monitors messages between agents to detect undesirable

behaviours such as failures. Collecting the information, the system generates global information of

interdependence between agents and expresses it in a graph. This interdependence graph enables us to detect

or predict undesirable behaviours. This paper also shows that the system can optimize performance of a

MAS and improve adaptively its reliability under complicated and dynamic environment by applying the

global information acquired from the interdependence graph to a replication system.

1 INTRODUCTION

A multiagent system (MAS) (Weiss, 1999) has

recently gained public attention as a method to solve

competition and cooperation in distributed systems.

However, MAS's vulnerability due to the

propagation of failures prevents it from being

applied to a large-scale system. It is indispensable

for constructing a ubiquitous network society to

raise MAS’s reliability on a large scale.

This paper proposes an approach to improve the

efficiency and reliability of distributed systems. The

approach monitors messages between agents to

detect undesirable behaviours such as failures.

Collecting and reserving the information, it

generates global information of interdependence

between agents and expresses it in a graph. This

interdependence graph enables us to detect or predict

undesirable behaviours.

Replication systems (Guerraoui, 1997) may

arrange multiple replicas with the same contents on

a network. They are regarded as an effective

technique for raising the fault tolerance of

distributed systems. Such systems have replicas

replaced with faulty agents in order to realize fault

tolerant applications. However, they are not suitable

for large-scale systems because replication cost may

increase in proportion to the number of replicas.

This paper also proposes an adaptive replication

system and replication policies that can efficiently

reduce replication cost and raise performance.

2 PROBLEMS OF PREVIOUS

MONITORING METHODS

Monitoring is usually adopted as a technique for

raising reliability of MAS that includes undesirable

behaviours. Some techniques are proposed and

problems of these techniques are shown below: An

immunity network (Ishida, 1996) depends on

knowledge agents have locally. It is a self-diagnostic

system, and it becomes very complicated because

fault tolerance and performance may be dependent

on network configuration. Moreover, a method of

Kaminka et al. (Kaminka, 2002) is based on a

recognition model of procedure to identify

disagreement of states. As the procedure is static and

premised on closed systems, this approach cannot be

adapted for changing environments. Finally, a

method of Horling et al. (Horling, 2001) is a

diagnostic system of distributed systems that uses

certain fixed failure models. It can optimize

system’s performance locally, but not globally. As a

result, this approach is not suitable for a MAS under

open environments where agents are added or

deleted on large scale and in real time.

199

Matsumoto K., Maruo T., Tanimoto A. and Mori N. (2008).

AN ADAPTIVE AND DEPENDABLE SYSTEM CONSIDERING INTERDEPENDENCE BETWEEN AGENTS.

In Proceedings of the Tenth International Conference on Enterprise Information Systems - SAIC, pages 199-202

DOI: 10.5220/0001669901990202

 SciTePress

3 MONITORING METHOD

USING INTERDEPENDENCE

GRAPH

3.1 Interdependence Graph

Collecting related information, the monitoring

system generates global information of

interdependence between agents and expresses it in a

graph. A domain agent is related to a node as shown

in Figure 1. This situation is expressed as labeled

graph (N, L, W). It is called an interdependence

graph. Where, N is a node set, L

i,j

is a link from

node N

to N

, and W

i,j

is a weight labeled to a link

i,j

. W

i,j

corresponds to importance of the

interdependence between agent i and j, and n is the

total number of nodes. If domain agents are added or

deleted, the graph is updated dynamically.

3.2 Monitoring Architecture

Monitoring task consists of two processes:

acquisition of the information for updating an

interdependence graph, and the graph analysis for

controlling domain agents. This information is

standard indicators like communication load and

processing time, agents’ characteristics and so on.

The architecture is shown in Figure 2. This

distributed observation mechanism corresponds to

the organization of agents that react in adaptation to

changing environments. There are two kinds of

agents: monitoring agents and a host monitoring

agent. A monitoring agent is associated with each

domain agent, and with the host monitoring agent.

The monitoring agent collects the individual

information as a monitor. Each monitoring agent

Figure 1: Interdependence graph.

Figure 2: Monitoring architecture.

communicates with the host monitor agent, and

transmits local information acquired by monitoring.

The host monitor agent makes it into global

information such as the total number of exchanged

information. The monitoring agent reflects various

changes of the domain agent in the interdependence

graph. Because agent environment changes in real

time, this graph is not static, and it is updated

dynamically if agents in the domain level are added

or deleted.

3.3 Updating Algorithm of Weights

Indexes for updating algorithm of weights are shown

below:

− monitoring time interval Δt

− communication load Q(Δt)

Q(Δt) = operator1(Q

(Δt), . . . , Qn,n(Δt)) (4)

− the number of transmitting messages NM(Δt)

Where Q

i,j

(Δt) and NM

i,j

(Δt) express respectively the

amount of communication data and the number of

transmitting messages from agent i to j in time

interval Δt, operator1 and operator2 are set operators

and they are usually average operators.

The outline of the updating algorithm that

updates the weights W

i,j

of the interdependence

graph is shown below.

1. Repeat the following for agent j (j ≠ i).

N = {Ni} i=1,…,n

(1)

L = {Li,j} i=1,…,n, j=1,…,n

(2)

W = {Wi,j} i=1,…,n, j=1,…,n

(3)

NM(Δt)=operator2(NM

(Δt), . . . ,NMn,n(Δt)) (5)

Domain

Level

Monitoring

Level

Domain

Agent

Domain

Agent

Domain

Agent

Domain

Agent

Monitoring

Agent

Monitoring

Agent

Monitoring

Agent

Monitoring

Agent

Host

Monitoring

Agent

Message

Control

Event

Interdependence Graph

0.3

0.8

0.3

0.1

0.5

0.7

0.2

0.4

0.9

0.6

0.3

ICEIS 2008 - International Conference on Enterprise Information Systems

200

2. Calculate formulas (6) and (7)

Qtemp= (Q

(Δt)−Q(Δt))/Q(Δt) (6)

NMtemp= (NM

i,j

(Δt)−NM(Δt))/NM(Δt) (7)

3. Updating weight W

i,j

(t+Δt)=W

i,j

(t)+α×Operator3(Qtemp,NMtemp) (8)

4. End of repetition

Parameter α is a discount rate in which a new

situation is reflected in the existing weight.

4 ADAPTIVE MULTIAGENT

ARCHITECTURE

4.1 Adaptive Replication System

Our adaptive multiagent architecture is shown in

Figure 3. This architecture consists of a monitoring

system and a replication server that manage domain

agents and replicas. In our adaptive replication

system, you can adjust dynamically the number of

replicas according to each agent's importance

obtained from the interdependence graph.

The replication system creates some replication

groups, and each group consists of one domain agent

and some replicas. The algorithm adjusts adaptively

the number of replicas in proportion to the

importance of a node.

Figure 3: Adaptive multiagent architecture.

4.2 An Algorithm to get the Number of

Replicas using Interdependence

The interdependence graph is useful for grasping the

influence of failures and the fault tolerance of

multiagent systems. The proposed monitoring

system estimates an agent's importance by executing

a set operation (operator4) on the weights W

i,j

that

are labeled to each agent's input-and-output links. It

is shown in formula (9).

= operator4 (W

j=1, ---, m) (9)

is the importance of agent i and m is the degree of

node i. The w

is used to compute the number of

replicas (rep

) for agent i in our adaptive replication

system. Formula (10) shows this relation.

rep

= round [r

+ r

max

×w

/W] (10)

is an initial number of replicas, r

max

is the

maximum value of the total number of replicas a

system designer sets up beforehand, and W is the

sum of all agents' w

5 SIMULATION EXPERIMENT

5.1 Virtual eMarket Multiagent System

The simulation is carried out for a virtual eMarket

multiagent system (eMarket MAS):

1. Intermediate-goods producers sell intermediate

goods to consumption-goods producers.

2. Consumption-goods producers process the

intermediate goods and make consumption

goods.

3. Consumers purchase the consumption goods

from consumption-goods producers.

A model of the two stage market used in the

simulation is shown in Figure 4. In this model, there

are two markets:

− Market A: Intermediate-goods market

− Market B: Consumption-goods market

Furthermore, agents constituting markets are

classified into the following four kinds of agents

according to their roles: Sellers, buyers,

buyer/sellers, and market management agents.

Sellers and buyers only sell intermediate goods or

purchase consumption goods respectively. The

buyer/sellers play different roles in each market.

They buy intermediate goods in Market A, and sell

consumption-goods in Market B. Finally, one

Monitoring System

lication Serve

Calculation of

Agent’s Importance

Update of

Interdependence

Monitoring

Replication

Mutual

Operation

Event

System

Even

Replication

Control

Monitoring

Level

Domain

Level

Agent

AN ADAPTIVE AND DEPENDABLE SYSTEM CONSIDERING INTERDEPENDENCE BETWEEN AGENTS

201

Figure 4: Two stage market model.

market management agent exists in each market, and

it controls registration or deletion of agents

constituting markets. The market management agent

mediates dealings. The markets deal with dealings

by auctions, and the scenario is described below:

1. A seller gives a market management agent a

reserve price.

2. The market management agent accepts bids

from buyers during a certain fixed time.

3. The market management agent cuts a deal with

the buyer that presented the highest bid price

exceeding the reserve price in a bid deadline.

4. If there is no corresponding bid, dealings are

abortive and the seller again presents the price

that is less than the previous one.

5.2 Experiment

Fault generators put a total of 100 agents to stop

within 10-minute simulations. If an agent exhausts

replicas prepared beforehand, the dealing scenario

cannot be completed. This means the simulation is a

failure. We count the number of successful

simulations when we change the total number of

replicas r

max

from 0 to 30. Simulations are performed

40 times in total. Parameters used for the experiment

are as follows: Market management agents: 2,

buyer/sellers: 50, buyers: 24, sellers: 24, and total

100 agents. Monitoring interval is 500ms. Discount

rate α is 1.0. Initial number of replicas r

is 1.

The experiment results are shown in Figure 5.

When r

max

is set up as 10, the rate of success reaches

80%. If r

max

is set as 20 and over, the success rate

becomes 100%. It proves that reliability of the

system is maintainable by setting r

max

as 20 and over

in this experiment environment, and it also means

that replication cost can be held down efficiently.

The parameter r

max

is important for systems and

must be set up suitably.

Figure 5: Relationship between r

max

and success rate.

6 CONCLUSIONS

To improve the efficiency and reliability of

multiagent systems, this paper has proposed an

algorithm

(i) to update the interdependence graph and

(ii) to adjust adaptively the number of replicas in

proportion to the importance of a node using the

interdependence graph.

It also has proposed an adaptive replication system

that uses global information acquired by monitoring

to improve fault tolerance of multiagent systems.

The method has been applied to an e-market MAS.

The simulation results show that the method can

optimize performance of a MAS and improve its

reliability by applying the global information

acquired from the interdependence graph to

replication systems.

REFERENCES

Weiss, G., 1999. Multiagent Systems -A Modern Approach

to Distributed Artificial Intelligence-. The MIT Press,

pp.79-120.

Guerraoui, R., Schiper, A., 1997. Software-based

replication for fault tolerance. IEEE Computer ,

Vol.30, No.4, pp.68-74.

Ishida, Y., 1996. An Immune Network Approach to

Sensor-based Diagnosis by Self-organization.

Complex Systems, Vol.10, No.1, pp.73-90.

Kaminka, G. A., Pynadath, D. V., Tambe, M, 2002.

Monitoring Teams by Overhearing: A Multi-agent

Plan-Recognition Approach. Intelligence Artificial

Research, Vol.17, pp.83-135.

Horling, B., Benyo, B., Lesser,V., 2001. Using Self-

Diagnosis to Adapt Organizational Structures. Proc. of

5th International Conference on Autonomous Agents,

pp.529-536.

Market A

Market B

Market

Management Agen

Market

Management Agent

Seller

Buyer

Buyer/Seller

Seller

Success rate

(

)

ICEIS 2008 - International Conference on Enterprise Information Systems

202