INFLUENCE OF NEIGHBORHOOD AND SELF

REORGANIZATION IN NETWORKED AGENTS

Udara C. Weerakoon and Vicki H. Allan

Utah State University, Logan, UT, U.S.A.

Keywords: Multi-agent system, Reorganization, Simulation, Coalition, Hedging environment.

Abstract: In a network graph in which nodes represent agents and edges represent "can work with" relationships,

coalitions form. Such coalitions satisfy the skill set requirements of a task while still obeying partner

requirements. Agents composing a coalition must form a connected subgraph in the network graph. There is

no centralized control, and agents are free to propose any coalition that satisfies both the skill set and partner

requirements. In this research, strengths of various coalition formation strategies are compared with respect

to both success and profit. To determine the quality of the solution and for comparison purposes, we

temporarily remove the restriction that an agent can belong to a single proposed coalition and that a task can

be proposed by a single coalition (i.e. hedging environment). In addition, agents are given the ability to

dynamically reorganize their partner connections in an attempt to improve utility. Agents employing

egalitarian, intelligent and inventory reorganization are compared with agents employing structural and

performance reorganization.

1 INTRODUCTION

We model the coalition formation problem as a

network graph in which nodes represent agents and

edges represent a "can work with" relationship. Each

agent possesses a single primary skill. Tasks require

a set of skills that must be present in the coalition for

the duration of task execution. Coalitions are

restricted to sets of agents linked via edges.

Reorganization is viewed as the mechanism

enabling individual agents to change their

connections dynamically without explicit external

commands (Marzo Serugendo et al., 2005). This

behavior can be generated in multi-agent systems in

several ways (Barton and Allan, 2008; Gaston and

Jardins, 2005; Thadakamalla et al., 2004). This

paper performs a comparative analysis of various

strategies of task selection and coalition formation.

Some strategies introduce specialist agents to the

organization (Hoogendoorn, 2007) to manage each

agent’s connections. Yet other methods, such as

organizational self-design (Kamboj, 2009), achieve

reorganization by dynamic spawning and merging

agents. In our model, we use autonomous agents to

improve and analyze reorganization.

2 RELATED WORK

In Abdallah and Lesser's work (Abdallah and Lesser,

2007), agents organize themselves in an overlay

network in which agents only interact with

neighbors. Similarly, in our method, agents

reorganize. However, Abdallah and Lesser restrict

their problem to that of task allocation (assigning

one agent to do a task) rather than coalition

formation. Gaston and Jardins (Gaston and Jardins,

2005) consider social networks and task formation

with multiple skills per task, but do not have varying

agent types.

In Barton and Allan's work (Barton and Allan,

2008), self-organized social networks under

changing resource requirements are considered.

Edges in the social network can be modified by

either adjacent agent. Such modification is termed

rewiring. However, the results lay at a low range of

efficiency and performance, typically less than 45%.

In our research, we extend these results by showing

that the efficiency/performance is often dictated by

the maximum connections each agent maintains.

376

Weerakoon U. and Allan V..

INFLUENCE OF NEIGHBORHOOD AND SELF REORGANIZATION IN NETWORKED AGENTS.

DOI: 10.5220/0003140703760379

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 376-379

ISBN: 978-989-8425-41-6

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

a) b)

Figure 1: Networked agent simulation: a graphical representation of the networked multi-agent system. Each node

represents an agent while an edge represents a relationship which indicates “can work with”. Nodes in blue are

uncommitted agents which are available to form a coalition or reorganize themselves in a better neighborhood. Nodes in red

are committed agents which are executing a task in a coalition (a). Nodes in green are committed agents in a partially

formed coalition for a given task. Edges in red are newly created relationships as a result of reorganization (b). Shaded

edges represent an abandoned edge due to reorganization.

3 SIMULATION

In this research, our goal is to determine strategies

which improve the success of a distributed coalition

formation network. Tasks have equal utility and are

generated at regular intervals during the simulation.

We define success-rate of an agent as the

fraction of successfully completed coalitions divided

by the total number of coalitions joined.

In our model, agents first consider joining

coalitions to which their neighbors belong (partner

requirement). Here, neighbor means a node

connected to an agent by an edge.

The state of an agent, based on the coalition

perspective, is as follows:

 active - executing a current task

 committed - has joined a coalition that has not

begun execution

 uncommitted - agent which can propose a new

coalition or join a coalition proposed by a neighbor

The set of agents that an agent can see in

following paths of a predefined length is termed

communication depth.

4 AGENT TYPES

4.1 Random Agents

Random agents are a primitive type of agent that

join coalitions (if possible) and otherwise propose.

No specific criterion is used to select the task to

join/propose.

4.2 Strategic Agents

Strategic agents select the coalition to join based on

a blend of (1) the coalition with the highest percent

of committed agents and (2) the coalition for which

peers (i.e. neighbors of neighbors) have the best

match with skills needed for the coalition. If the

agent is not satisfied with choices for joining a

coalition, the strategic agent can then propose a new

coalition, selected based on whether its peers have a

sufficiently high chance of satisfying the necessary

criteria. As a last resort, the strategic agent can

randomly propose a coalition.

5 REORGANIZATION

An agent has the ability to remove an edge between

INFLUENCE OF NEIGHBORHOOD AND SELF REORGANIZATION IN NETWORKED AGENTS

377

it and its neighbor and create a new edge with an

agent that is not a current neighbor. This is termed

reorganization. We study five types of

reorganization: performance reorganization

(Gaston and Jardins, 2005), structural

reorganization (Thadakamalla et al., 2004),

egalitarian reorganization, inventory

reorganization, and intelligent reorganization.

Table 1 summarizes the parameters of each

reorganization type in the connected agent network.

Table 1: Summary of each reorganization strategy.

Reorg. Trigger How Selected Can

refuse

Performance prob 1/|A| performance no

Structural prob 1/|A| most connections no

Egalitarian prob 1/|A| fewest

connections

Inventory prob 1/|A| needed skill no

Intelligent poor perfor-

mance

current skill

demand

yes

6 ENVIRONMENT

6.1 Hedging Environment

In our model, we compare the effects of allowing

agents to commit to multiple coalitions (that are not

yet executing), and we allow multiple possible

coalitions to be associated with the same task. The

tradeoff in this environment is the balance between a

higher number of successful coalitions and the cost

of discarding unsuccessful coalitions.

7 EMPIRICAL EVALUATION

In our first set of experiments (Figures 2-3), each

random, strategic or hedging agent is connected with

the same number of neighbors (number of

connections). No reorganization is done here. In our

second set of experiments (Figures 4-5), the

behavior of hedging agents employing five different

reorganizations are analyzed.

Consider Figure 2. We use the term saturation

point to indicate the point at which an agent achieves

a .9 (90%) performance-rate (in the tests). In the

hedging environment, agents reach the saturation

point at 10 connections. Strategic agents require 18

connections to achieve their saturation point.

Random agents require 20 connections to achieve

their saturation point.

Figure 3 shows the corresponding profit earned

by each agent simulation. Interestingly, in the

hedging environment, the profit degrades from its

maximum value when the number of connections is

more than 34, due to the higher communication cost

and insignificant improvement in reward.

Figure 2: Performance-rate Vs Agent Connections.

Figure 3: Profit Vs Agent Connections.

Agents employing reorganization give us the

opportunity to understand the effects on the

performance-rate, and interestingly, the dilemma of

society, a tragedy of commons (Axelrod, 1997).

Figure 4 depicts the performance-rate of the

hedging environment with five types of

reorganization. Reorganization increases the

performance-rate even in the hedging environment.

It is of note that with hedging, egalitarian

outperforms the others. Agents employing

performance and inventory reorganizations diminish

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

378

their maximum value at the time points of 2800 and

3500, due to the number of isolated agents (agents

without neighbors/connections).

Figure 5 depicts the number of isolated agents in

each agent simulation due to reorganization.

Figure 4: Performance-rate Vs Time (with reorganization).

Figure 5: Isolated Agents Vs Time (with reorganization).

8 CONCLUSIONS

Our research shows that strategic agents are

significantly better than random agents and, for a

high number of connections, are competitive with

the upper bound (hedging environment).

Success of local strategies (without hedging)

depends heavily on having sufficient neighbors. The

results show us that hedging agents are capable of

earning more profit than others as their increased

success negates the extra cost of discarding

unnecessary coalitions, and strategic agents are

competitive with them.

Agents employing egalitarian reorganization

outperform all other reorganizations. Performance

and inventory reorganization result in a high number

of isolated agents.

A better plan would be to have a mixture of

strategies: some which directly pursue goals and

others which seek to rebuild and utilize agents which

have been abandoned in the simulation.

ACKNOWLEDGEMENTS

This material is based in part upon work supported

by the National Science Foundation under grant

number 0812039

REFERENCES

Abdallah, S. and Lesser, V., 2007. Multiagent

reinforcement learning and self-organization in a

network of agents. In AAMAS (Autonomous agents

and multiagent systems). Honolulu, Hawaii.

Axelrod, R. 1997. Building New Political Actors. The

Complexity of Cooperation: Agent-based Models of

Competition and Collaboration. Princeton University

Press.

Barton, L. and Allan, V. H., 2008. Adapting to Changing

Resource Requirements for Coalition Formation in

Self-Organized Social Networks. In IAT (International

Conference on Intelligent Agent Technology). Sydney,

Australia, December 2008.

Gaston, M. E. and Jardins, M. Des, 2005. Agent-

Organized Networks for Dynamic Team Formation. In

AAMAS (Autonomous Agents and Multi-Agent

Systems). Utrecht, Netherlands, pp. 230-237.

Hoogendoorn, M., 2007. Adaptation of organizational

models for multi-agent systems based on max flow

networks. In IJCAI (International Joint Conference on

Artificial Intelligence), India, pp. 1321-1326.

Kamboj, S., 2009. Analyzing the tradeoffs between

breakup and cloning in the context of organizational

self-design. In AAMAS (Autonomous Agents and

Multi-Agent Systems). Hungary, pp. 829-836.

Marzo Serugendo, G. Di, Gleizes, M. P. and Karageorgos,

A., 2005. Self-organization in multi-agent systems.

The Knowledge Engineering Review, 20(2), pp. 165–

189.

Thadakamalla, H. P., Raghaven, U. N. K., and Albert, R.,

2004. Survivability of Multiagent-based Supply

Networks: A Topological Perspective. IEEE

Intelligent Systems, vol.19, pp. 24-31.

INFLUENCE OF NEIGHBORHOOD AND SELF REORGANIZATION IN NETWORKED AGENTS

379