COORDINATION IN OPEN AND UNSTRUCTURED INTELLIGENT

AGENT SOCIETIES

Using Distributed Planners on Top of a Semantic Overlay Network

Ant

onio Lu

ıs Lopes and Lu

ıs Miguel Botelho

Instituto das Telecomunicac¸

oes, Av. Rovisco Pais, 1, Lisboa 1049-001, Portugal

Keywords:

Multi-agent coordination, Planning algorithms, Graphplan, Peer-to-peer networks, Semantic overlay net-

works.

Abstract:

Collaborative environments, where multiple heterogeneous agents (managing several resources) can coop-

erate in pursuing common and individual goals, are a step forward in creating real-world agent societies.

However, current research in agent negotiation and in service coordination is still not enough for building

such an agent-based society, capable of jointly solving complex planning problems and still achieve overall

good performance. Most often, current work relies on either some centralised component or pre-deﬁned so-

cial structure, which can compromise the system in terms of scalability, openness and robustness, and fails to

address general problems. By using efﬁcient network search algorithms and network evolution techniques it

is possible to build and maintain a semantic overlay network from a totally unstructured distributed network,

which in turn will simplify and optimize the distributed planning process amongst heterogeneous agents. We

developed distributed versions of well-known planners that operate on top of the referred semantic overlay

network and through a set of tests (using different scenarios) we were able to determine which is the best

algorithm.

1 INTRODUCTION

Our main goal is to develop a robust agent architec-

ture that enables agents to freely participate in dis-

tributed problem solving in open unstructured agent

societies. In such societies, agents receiving requests

from other agents are capable of using their own ca-

pabilities to handle the part of the problem for which

they have competence and resources, and distribute

the partially solved problem to other agents that can

possibly provide further contributions. This work

combines artiﬁcial intelligence, distributed problem

solving and peer-to-peer computing to address its two

main challenges: allow agents to partially contribute

to complex problems and to efﬁciently delegate un-

solved sub problems to other agents over unstructured

decentralised networks.

Cooperating in distributed problem solving re-

quires agents to be able to discover other agents to

delegate the sub problems for which they cannot con-

tribute. In previous research (Lopes and Botelho,

2008) we have shown that it is possible to improve

the resource coordination process in multi agent based

peer-to-peer networks by building, maintaining and

using a powerful semantic overlay network (Cre-

spo and Garcia-Molina, 2005) that is dynamically

learnt and updated by the discovery mechanism it-

self. The discovery and self-organisation process, in

which agents establish semantic connections amongst

them, thus fuelling the semantic overlay network, is

ﬁrst carried out by using efﬁcient and robust search

mechanisms and network evolution techniques (see

(Lopes and Botelho, 2008) for details). Once the se-

mantic overlay network is built, besides using it to

easily locate resources, agents can use it to perform

more complex tasks such as planning a solution for a

speciﬁc problem.

In this paper we focus on the planning stage, by

analysing and presenting a set of distributed planning

algorithms that can be used, on top of the semantic

overlay network, to solve planning problems in coop-

erative environments. Section 2 introduces the prob-

lem and presents related work. Section 3 describes

our approach and in section 4 we discuss tests and re-

sults. Finally, section 5 concludes the paper.

347

Luís Lopes A. and Miguel Botelho L. (2010).

COORDINATION IN OPEN AND UNSTRUCTURED INTELLIGENT AGENT SOCIETIES - Using Distributed Planners on Top of a Semantic Overlay

Network.

In Proceedings of the 2nd International Conference on Agents and Artiﬁcial Intelligence - Agents, pages 347-350

DOI: 10.5220/0002761403470350

 SciTePress

2 MOTIVATION

A typical planner has been seen as a producer of ac-

tion sequences that requires the following inputs: (i)

an initial state of the world; (ii) a description of a

goal state; and (iii) a set of possible actions, which

can be used in the generated plan to lead the agent

from the initial state to the goal state. As the need

for more processing power (for increasingly complex

problems) arose and the inadequacy of centralised ap-

proaches became evident, researchers turned to Dis-

tributed Problem Solving. However, taking advantage

of the decentralised control of such distributed envi-

ronments requires that coordination mechanisms exist

that are able to avoid conﬂicts that arise from the con-

current interactions of agents, which otherwise would

result in a turmoil.

The coordination process can be undertaken at dif-

ferent times and situations, depending on what is suit-

able for the speciﬁc domain to which it applies. In

spite of the variety of approaches, domains and con-

texts, we have concluded that they have the same

kind of limitations. Whether they rely on some sort

of centralised component or on a pre-deﬁned struc-

ture/knowledge that rules the activity of all entities

in the environment, these approaches show signs of

compromised scalability and robustness.

Most task reﬁnement and task allocation ap-

proaches rely on centralised components, such as

blackboards (Wellman, 1993), tables of capabilities

(Fung and Chen, 2005) or broker like auctioneers

(Wellman et al., 2001). An obvious limitation of these

approaches in very large networks is its non scalabil-

ity. Other approaches (de Weerdt et al., 2007) propose

the use of social structures to govern the task alloca-

tion process, but again these have to be provided a

priori by some human user.

The same limitations apply to planning coordina-

tion. Whether coordination is performed before, dur-

ing or after planning, most approaches rely on some

pre deﬁned structure or some centralised component

to govern the activity of the agents in the environ-

ment. Social laws (Shoham and Tennenholtz, 1992)

and other organizational based approaches (Abdallah

and Lesser, 2004) (Gaston and Desjardins, 2005) are

examples of systems where sets of rules or organiza-

tional structures are imposed to the agents societies.

Alternative approaches (Cox et al., 2003) use fully

connected networks to allow agents to know which

and when agents should be contacted. However, this

kind of network topology is very difﬁcult to manage

in large dynamic networks where high churn rates (the

rate at which agents enter or leave a network) make

the approach non scalable. Centralisation is often the

choice for most domains, whether it is by allowing

agents to coordinate themselves through an advertis-

ing blackboard (de Weerdt and Van Der Krogt, 2002),

or by using specialised central agents to perform post-

planning coordination (Cox and Durfee, 2005).

3 TECHNICAL APPROACH

By using a matchmaking process that analyses each

agents operators and establishes links between in-

puts/outputs and pre-conditions/effects of the opera-

tions that can be done by other agents (see details in

(Lopes and Botelho, 2008)), we dynamically build a

powerful Semantic Overlay Network. This abstract

layer subsumes the discovery process of the coordi-

nation environment by helping agents ﬁnd each other

based on their semantic relationships. However, for

an agent to create a plan to solve a speciﬁc problem,

it still needs a planning algorithm that is capable of

dealing with partial knowledge of the domain opera-

tors and deriving the potential contributions that can

be done by the agent to the referred problem. Our

work focuses on ﬁnding the appropriate planning al-

gorithm for this distributed environment.

We have studied several algorithms and decided

to use the well-known Graphplan planner (Blum and

Furst, 1997) because it is one of the best planning

algorithms and has been used as the basis for other

even more efﬁcient planning algorithms. We have

created different versions of the planner so we could

test its application to different distributed problems.

In section 3.1, we present our distributed version

of the Graphplan algorithm. Section 3.2 describes

an alternative version of the distributed Graphplan

algorithm that previously builds an operators-graph

through means-ends analysis.

3.1 Distributed Graphplan

In a centralised version of the Graphplan algorithm

an agent has full knowledge of the available opera-

tors. However, in a distributed setting, each agent may

only have knowledge of its own operators and of those

of selected neighbours or, in the case of the Seman-

tic Overlay Network, of agents that are semantically-

linked to it. So, we modiﬁed the algorithm to allow

partial contributions to the creation of the planning-

graph. The process is carried out as follows:

• The agent receiving a request (which includes a

description of the initial state and the goal state)

will analyse each level and determine which of

the operators (that it knows) can contribute to the

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

348

current planning-graph (if the agent is the ﬁrst re-

ceiving the request, it has to create the ﬁrst propo-

sition level, which will include all propositions of

the initial state) and adds them accordingly;

• The agent further analyses open propositions (for

which it was unable to contribute) and determines

(using the Semantic Overlay Network) which are

the most adequate agents to receive this partially-

complete planning-graph;

• Each agent receiving the planning-graph will ex-

ecute these same steps up to a point where a level

in the graph is reached where all goal propositions

exist and none of which are mutex;

• Once a planning-graph reaches this desired stage,

the agent holding the planning-graph at that time

executes the backward search that will ﬁnd a so-

lution plan – the agent can also request the assis-

tance of other agents in the backward search, but

each agent will use a different heuristic in the pro-

cess;

The termination of this process in a distributed en-

vironment is not trivial. In the centralised version

an agent can simply rely on the level-off property of

Graphplan

to conclude that the problem is impossi-

ble to solve. For an agent with only partial knowledge

of the world, it is impossible to know if a levelled-off

graph means that the problem is impossible or that

the agent simply does not have enough knowledge to

complete it. This could lead to a near-inﬁnite pro-

cess of forwarding partially solved problems between

agents. To avoid this situation, we use a similar mech-

anism as the one used in peer-to-peer search algo-

rithms, where a time-to-live setting is used to specify

the allowed number of forwards that can be done with

a single request.

3.2 Distributed Graphplan with

Operators-Graph

In most domains, the size of the goal state is a lot

smaller than the initial state, which means that some

of the propositions contained in the initial state may

be completely irrelevant to reach the goal state. As

most forward-based planners, Graphplan suffers from

the problem of distraction, where the planner consid-

ers all propositions in the initial state even if they will

not help reach a solution plan. In fact, these unnec-

essary propositions can be very time-consuming, thus

degrading the performance of the planner. Therefore,

they should be avoided.

When two subsequent levels are equal and non-mutex

goal propositions have been reached yet

To cope with this problem, we have used a sim-

ilar approach to (Kambhampati et al., 1997). We

use means-ends analysis in the Graphplan algorithm,

by ﬁrst producing an operators-graph (Smith and

Peot, 1996) using a backward-chaining process start-

ing from the goal state. Since it only considers the

propositions in the goal state, the operators-graph

will produce a graph with only relevant actions.

This planner uses a similar process to the one used

in the generation of the planning-graph but in a differ-

ent direction. It ﬁnds operators that can contribute to

propositions in the goal state and the pre-conditions

of those operators become the new goal propositions.

The planner proceeds with this process until it reaches

a level in which all propositions are contained in the

initial state. After the graph has been generated, the

normal forward generation of the planning-graph can

take place, except this time it will only consider the

operators that are contained in the operators-graph,

thus signiﬁcantly reducing the size of the graph and

the number of calculations. The drawback is, obvi-

ously, the overhead introduced by the generation of

the operators-graph.

4 TESTS AND RESULTS

In order to test the distributed Graphplan and the

operators-graph-based version of Graphplan algo-

rithms, we deployed two different testing scenarios

but due to space limitations we will only show the re-

sults for the Rescue Agents scenario. In this scenario,

agents represent entities that participate in a rescue

operation after the occurrence of a natural disaster,

where they have to perform operations such as clear-

ing roads, putting out ﬁres and providing assistance to

injured people. This scenario is characterized as hav-

ing a small number of different entities but with a high

degree of complexity due to the high level of interac-

tion/cooperation that is needed between the agents.

We have performed a set of preliminary tests for

this scenario using increasingly complex variations

(different number of injured people) on both algo-

rithms. As depicted in Figure 1, the results show that

the operators-graph-based version of the distributed

Graphplan algorithm is more scalable than the dis-

tributed Graphplan algorithm. This is strongly linked

to the fact that for more complex or large problems,

the means-ends analysis is effective in reducing the

search scope of the planner, in spite of the introduced

overhead.

COORDINATION IN OPEN AND UNSTRUCTURED INTELLIGENT AGENT SOCIETIES - Using Distributed Planners

on Top of a Semantic Overlay Network

349

Figure 1: Comparison between Graphplan and Operator-

graph-based Graphplan in the Rescue Agents scenario

5 CONCLUSIONS AND FUTURE

WORK

In this paper we have described the approach taken in

a cooperative planning environment, where we have

deployed a distributed network of problem solving

agents by using a Semantic Overlay Network and dis-

tributed Graphplan-based algorithms. Preliminary re-

sults show that the approach can be considered scal-

able and efﬁcient. The results are still preliminary and

we intend to perform a more thorough analysis of the

testing scenarios and include other scenarios to ad-

dress the robustness of the approach in a large variety

of problems.

ACKNOWLEDGEMENTS

This work has been supported in part by the Por-

tuguese Foundation for Science and Technology un-

der the scholarship grant SFRH/BD/27533/2006. The

authors would also like to thank the support of

the Lisbon University Institute and the Instituto de

Telecomunicac¸

oes (IT).

REFERENCES

Abdallah, S. and Lesser, V. (2004). Organization-based co-

operative coalition formation. In Proceedings of the

IEEE/WIC/ACM Int. Conf. on Intelligent Agent Tech-

nology, pages 162–168.

Blum, A. and Furst, M. (1997). Fast planning through

planning graph analysis. Artiﬁcial intelligence, 90(1-

2):281–300.

Cox, J. and Durfee, E. (2005). An efﬁcient algorithm for

multiagent plan coordination. In Proceedings of the

4th Int. Joint Conf. on Autonomous Agents and Multi-

agent Systems, pages 828–835. ACM New York, NY,

USA.

Cox, M., Elahi, M., and Cleereman, K. (2003). A

distributed planning approach using multiagent goal

transformations. In Proceedings of the 14th Midwest

Artiﬁcial Intelligence and Cognitive Science Confer-

ence, pages 18–23.

Crespo, A. and Garcia-Molina, H. (2005). Semantic overlay

networks for p2p systems. In Proceedings of the 3rd

Int. Workshop on Agents and Peer-to-Peer Computing,

page 1, New York, NY, USA. Springer.

de Weerdt, M. and Van Der Krogt, R. (2002). A method

to integrate planning and coordination. Planning with

and for Multi-Agent Systems, pages 83–88.

de Weerdt, M., Zhang, Y., and Klos, T. (2007). Distributed

task allocation in social networks. In Proceedings of

the 6th Int. Joint Conf. on Autonomous Agents and

Multiagent Systems. ACM New York, NY, USA.

Fung, R. and Chen, T. (2005). A multiagent supply

chain planning and coordination architecture. Interna-

tional Journal of Advanced Manufacturing Technol-

ogy, 25(7–8):811–819.

Gaston, M. and Desjardins, M. (2005). Agent-organized

networks for dynamic team formation. In Proceed-

ings of the 4th Int. Joint Conf. on Autonomous agents

and multiagent systemsnternational joint conference

on Autonomous Agents and Multiagent Systems, pages

230–237. ACM New York, NY, USA.

Kambhampati, S., Parker, E., and Lambrecht, E. (1997).

Understanding and extending graphplan. In Pro-

ceedings of the 4th European Conference on Plan-

ning: Recent Advances in AI Planning, pages 260–

272. Springer-Verlag London, UK.

Lopes, A. and Botelho, L. (2008). Improving multi-agent

based resource coordination in peer-to- peer networks.

Journal of Networks, 3(2):38–47.

Shoham, Y. and Tennenholtz, M. (1992). On the synthesis

of useful social laws for artiﬁcial agent societies. In

Proceedings of the National Conference on Artiﬁcial

Intelligence, pages 276–276.

Smith, D. and Peot, M. (1996). Suspending recursion in

causal-link planning. In Proceedings of the 3rd Int.

Conf. on Artiﬁcial Intelligence Planning Systems.

Wellman, M. (1993). A market-oriented programming envi-

ronment and its application to distributed multi com-

modity ﬂow problems. Journal of Artiﬁcial Intelli-

gence Research, 1(1):1–23.

Wellman, M., Walsh, W., Wurman, P., and MacKie-

Mason, J. (2001). Auction protocols for decentralized

scheduling. Games and Economic Behavior, 35(1-

2):271–303.

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

350