Artificial

Intelligence
Artificial Intelligence 94 (1997) 79-97

Negotiation and cooperation in

multi-agent environments *
Sarit Kraus a,bJ
a Department of Mathematics and Computer Science, Bar Ilan University Ramat Can, 52900 Israel
b hstith!te for Advanced Computer Studies, University ofMaryland, College Park, MD 20742, USA

Abstract
Automatic intel~gent agents i~abiting a shared environment must coordinate their activities.
Cooperation-not
merely coordination-may
improve the performance of the individual agents or
the overall behavior of the system they form. Research in Distributed Artificial Intelligence (DAI)
addresses the problem of designing automated intelligent systems which interact effectively. DA1
is not the only field to take on the challenge of understanding cooperation and coordination. There
are a variety of other multi-entity environments in which the entities coordinate their activity and
cooperate. Among them are groups of people, animals, particles, and computers. We argue that
in order to address the challenge of building coo~nated and collabom~d intelligent agents, it is
beneficial to combine AI techniques with methods and techniques from a range of muIti-entity
fields, such as game theory, operations research, physics and philosophy. To support this claim, we
describe some of our projects, where we have successfully taken an interdisciplinary approach.
We demomtrate the benefits in applying multi-entity methodologies and show the adaptations,
modifications and extensions necessary for solving the DA1 problems. @ 1997 Elsevier Science
B.V.
~ey~ur~~~ Dis~buted Artificial Intelligence; Multi-agent systems; Cooperation; Negotiation

1. Introduction
One of the greatest challenges for computer science is building computer systems that
can work together. The integration of automated systems has always been a challenge,
*This is an extended version of a lecture presented upon receipt of the Computers and Thought Award at
the 14th International Joint Conference on Artificial Intelligence in Montreal, Canada, August 1995.
Email: [email protected] or [email protected].
0004-3702/97/$17.00 @ 1997 Elsevier Science B.V. All rights reserved.
PIISOOO4-3702(97)00025-S

80

S. Kraus/Artijicial

Intelligence

94 (1997) 79-97

but as computers have become more sophisticated, the demands for coordination and
cooperation have become more critical. It is not only basic level components such as
printers, disks, and CPUs, but also high-level complex systems that need to coordinate
and cooperate.
Examples of such intelligent systems include:
l automated
agents that monitor electricity transformation networks [ 321;
l teams of robotic
systems acting in hostile environments
[5];
l computational
agents that facilitate distributed design and engineering [ 541;
l distributed
transportation and planning systems [ 25,561;
l intelligent
agents that negotiate over meeting scheduling options on behalf of people
for whom they work [67]; and
l Internet
agents that collaborate to provide updated information to their users.
In these environments, even when coordination is not required, cooperation may improve
the performance of the individual agents or the overall behavior of the system they
form.
Problems of coordination and cooperation are not unique to computer systems, but
exist at multiple levels of activity in a wide range of populations. People pursue their
own goals through communication
and cooperation with other people or machines. Animals interact (with limited language), cooperate with each other, and form communities.
Particles interact with each other and compose different types of material and phases
of matter. Although most computers currently act in multicomputer environments,
the
interaction among them is generally restricted, and they interact under strict rules. Negotiation or other sophisticated interactions rarely occur among computers. In general,
the levels of negotiation, bidding, voting, and other sophisticated interactions that characterize natural coordinating systems are absent.
Recent research in Distributed Artificial Intelligence
(DAI) aims to increase the
power, efficiency, and flexibility of intelligent automated systems (agents) by developing sophisticated techniques for communication
and cooperation among them. In my
research, I have addressed the challenge of building coordinated and collaborated intelligent agents by combining AI techniques with methods and techniques from various
fields that study multi-entity behavior.
I argue that an interdisciplinary
approach is beneficial for the development of coordinated and cooperative intelligent agents. Because these fields, which study multi-entity
behavior, are not concerned with agent design, one might think that they are not relevant
for DAI. Our experience is quite the contrary. It is true that these fields do not solve AI
problems, but they have thought about a wide range of issues that are important to the
design of intelligent agents, and they provide techniques, sometimes with proven properties or methods for proving properties that are useful to adopt for designing agents. DAI
researchers still have a lot of work left in order to adapt these methods for their needs;
however, they do not need to start from scratch. In this paper, we show by example the
advantages and the challenges of building on other work.
The amount of work done in the related fields is overwhelming.
Thus, a major
challenge in taking an interdisciplinary
approach is determining which technique to use.
There are several parameters that influence the choice of the appropriate techniques for
a DAI application:

S. Kraus/Artificial Intelligence 94 (1997) 79-97

81

(1) Lie level of cooperation among the agents: cooperative agents which work toward satisfying the same goal versus agents which are self-motivate and try to
maximize their own benefits. There are intermediary cases where self-motivated
agents join together to work toward a joint goal.
(2) Regulations and protocols: environments where the designers of the agents can
agree on regulations and protocols for the agents interaction versus situations
with no pre-defined regulations and protocols.
(3) Number of agents: a very large number of agents (hundred or more) versus a
few agents which communicate and coordinate their actions.
(4) Type of agents: systems of automated agents versus systems composed of people
and automated agents.
(5) C~~~~ni~at~on and computation costs: the availability and cost of communication among the agents and their computation capabilities and costs.
Any DA1 task can be characterized according to these dimensions. This characterization guides the choice of the multi-entity technique that can be applied to the specific
task.
Consider the deveIopment of automated agents for buying and selling items on the
Web, such as clothes and furniture. Suppose there are several enterprises, each with
several kinds of goods which they sell to users or to other enterprises. Each enterprise
has intelligent seller and buyer agents. The job of the seller agent is to sell the enterprises
goods to other enterprises through their buyer agents or to users. The job of a buyer
agent is to obtain from other enterprises the goods that are missing ffom the stock of
its enterprise. Several different DA1 problems may arise in such a fr~ework:
in the interaction between two automated agents belonging to different enterprises, the agents are self-motivated, but may benefit from cooperation. The
designers of the agents may agree upon regulations for the interaction, the number of agents of each interaction is limited, and they can communicate and have
computation capabilities.
A seller agent of an enterprise may try to sell some goods to a person. In this
case, the person will prefer a non-structured interaction, and it is more difficult
to set regulations and protocols for the interaction in advance.
Two agents of the same enterprise may work together toward the same goal:
increasing the benefits of their enterprise. In this case, the agents are cooperative,
regulations and protocols can be set in advance, the number of agents is limited,
they are automated, and they can communicate.
In each of these three cases, there is a different multi-entity technique that should be
applied.
In this paper, we will examine different DAI tasks and will discuss the application of
game-theoretic techniques (Section 2), physics models (Section 3), operations research
methods (Section 4)) and informal models of cooperation and coordination (Section 5)
to DAI environments.
2Research in DAI is divided into two basic classes: distributes Problem Solving (DPS) and belts-Agent
Systems (M.4) [ 61. Cooperative agents belong to the DPS class, while self-motivated agents belong to the
MA class.

S. Kraus/Art@ial

82

2. The application

of game-theoretic

Intelligence 94 (1997) 79-97

techniques to multi-agent

environments

Researchers in DA1 have considered problems related to task allocation and resource
sharing where the agents are self-motivated, as in the following examples:
o situations where airplanes belonging to different airlines need to share the limited
resources of the same airport, and it is necessary to find a mechanism that will give
priority to planes with less fuel on board [ 611;
o an electronic market populated with automated agents which represent different
enterprises and buy and sell (e.g., [8,17,74]);
o transportation centers that deliver packages and may cooperate to reduce expenses

[641;
o info~ation

servers that form coalitions for answering queries [ 361; and

o intelligent agents that negotiate over meeting scheduling options on behalf of people
for whom they work [67].
Using the five criteria presented in the Introduction
to characterize these examples,
we observe that in these examples the agents are self-motivated and try to maximize
their own benefits. The designers of the agents may agree in advance on regulations
and protocols for the agents interaction, In each interaction the number of agents is
usually small (less than a dozen agents) ; there are only automated agents which can
communicate and have computational capabilities.
In such situations we recommend the application of game-theoretic techniques. Game
theory studies mathematical models of conflict and cooperation between people. The
models of game theory are highly abstract representations of classes of real life situations
that involve individuals who have different goals or preferences [49] _ The active entity
in all game-theoretic
models is a player. Game-theoretic
models are divided into two
main types: noncooperative
models, in which the sets of possible actions of individual
players are primitives, and cooperative models, in which the sets of possible joint
actions of groups of players are primitive [53].
The abstract models of game theory can be used as a basis for the agents interaction protocols, when the designers of the agents agree to use them. Automated
agents in a DAI framework can be modeled by players of game-theoretic
models.
Since it is assumed in game theory that the players are self-motivated,
so should
the agents be in the environments
where these techniques are applied. DA1 addresses
both situations in which each individual agent acts by itself (e.g., [17,61,67]),
thus
calling for the application of noncooperative
models, and situations in which tasks
require that groups of agents work together (e.g., [ 36,641)) thus calling for cooperative models. In the first case, game-theoretic
models are appropriate when there
is only a handful of agents, and, in the second case, game theory may be applied
to a few dozens of agents. The use of game-theoretic
techniques requires substantial computations,
and, communication
capabilities are also usually needed. The examples described in the beginning of the section, as well as other situations that
satisfy these criteria, are cases where applying game-theoretic
techniques by DA1
researchers should be considered. We have applied both noncooperative
and cooperative gee-th~retic
models to DA1 situations. We briefly describe these attempts
here.

S. Kraus/Ar@cial

2.1. Use

Intelligence

94 (1997)

79-97

83

ofa strategic model of negotiation for resource sharing and task distribution

We first consider situations where a small number of self-motivated agents need to

share resources or can benefit from task distribution. In these situations, inter-agent
cooperation can be enhanced using negotiation strategies that enable agents to communicate their respective desires and to compromise in order to reach mutually beneficial
agreements.
The game-theoretic strategic approach to the bargaining problem3 provides a useful
foundation for designing such capabilities in systems. In this approach, agents negotiating maneuvers are moves in a noncooperative game, and the rationality assumption of
the negotiators [ 301 is expressed by using the Nash Equilibrium concept. 4 Our main
goal in this research was to define an acceptable protocol for the interactions among the
agents and to identify strategies for the agents p~icipating
in the negotiation. These
methods are applicable in the following example.
Example 1 (Data allocation in multi-agent environments). There is a set of several
(more than two) information
servers in the environment
which are connected by a
communication
network. Each server is located in a different g~graphical
area and
receives queries from clients in its area. In response to a clients query, a server sends
back information stored locally or information stored in another server, which it retrieves
from that server. 5 The information is clustered in datasets. 6
When a set of new datasets arrives, each new dataset has to be allocated to one
of the servers by mutual agreement among all of them. However, each server has its
own interests and wants to maximize its own utility, and thus the servers may be in
conflict concerning where to locate the new datasets. Furthermore, the servers have no
common interest and no central controller which can be used to resolve such conflicts.
We propose that these conflicts will be resolved via negotiations. In particular, we
propose a strategic negotiation model that takes into account the passage of time during
the negotiation process itself in order to solve this problem.
Our negotiation protocol is a process that may include several iterations. We assume
that servers can take actions in the negotiation only at certain times in the set Time =
(0, I, 2,. . .} that are fixed in advance and ordered (randomly). In each period t E Time,
if the neg~~tiation has not terminated earlier, a server whose turn it is to make an offer
at time t, will suggest a possible allocation for all the datasets considered, and each of
the other servers may either accept the offer or reject it or opt out of the negotiation.
If an offer is accepted by all the servers, then the negotiation ends, and this offer is
implemented.
If at least one of the servers opts out, then the negotiation ends. If no
server has opted out, but at least one of the servers has rejected the offer, the negotiation
3 Introductory books on game theory that discuss approaches to bargaining include [21,47,49,53].
4A pair of strategies (CT,T) is a Nash Equilibrium if, given 7, no strategy of Agent 1 results in an outcome
that Agent 1 prefers to the outcome generated by (CT,T), and similarly for Agent 2, given (T.
5A specific example of such a distributed knowledge system is the Data and information System component
of the Earth Observing System (EOSDJS) of NASA [50]. It is a distributed system which supports the
archiving and dis~bution of data at multiple and inde~ndent data centers.
4 A dataset corresponds to a clusrer in information retrieval, and to afile in the file allocation problem.

84

S. Kraus/Artijkial Intelligence 94 (1997) 79-97

proceeds to period t + 1, and the next server makes a counter-offer, the other servers
respond, and so on.
Using this negotiation mechanism, we showed that the servers have simple and stable
negotiation strategies that result in efficient agreements without delays. We have proved
that our methods yield better results than the static allocation policy currently used for
data allocation for servers in distributed systems.
The main question is, in general, what is the advantage of using game-theoretic
models for such problems, and what must be done in order to adapt them to DAI
environments. The strategic bargaining theory provides general frameworks for modeling
negotiation, but to apply them to the design of agents, we needed to address five
problems: choosing a strategic bargaining model which is applicable for the specific
DAI problem; matching the DAI scenarios with the game-theoretic definitions of the
chosen model; identifying equilibrium strategies; developing low complexity techniques
for searching for appropriate strategies; and providing utility functions.
For example, for the data allocation problem described in Example 1, we have chosen
Rubinsteins model of Alternative Offers [ 621. 7 The main property of this model is that
it takes into consideration the passage of time during the negotiation. This is useful for
environments of Example 1 since for a server participating in the negotiation process, the
time when an agreement is reached is very important. The model of Alternative Offers
provides formal definitions of players, possible agreements, the protocol of alternative
offers, and the notion of strategies. In order to apply these concepts to the data allocation
problem, we had to match the world state and formal definitions and modify them. For
example, in the data allocation scenario, a player is a server and an agreement is a
distribution of datasets to information servers.
Game theory proposes different notions of equilibria that capture different aspects of
stability. Given specific assumptions about the environments, game theory researchers
identify strategies that are in equilibrium. In order to address the third need mentioned
above, when applying game-theoretic techniques to DA1 environments, we formalized
the assumptions that are appropriate for our environments. For example, in the data
allocation scenario, all agents sustain a loss over time, there is a finite (but large) set of
agreements, and there are some agreements which are better for all agents than opting
out of the negotiations. In most of the cases, these assumptions are different from the
assumptions that are considered in game theory, and therefore we needed to identify the
equilibrium strategies under the DA1 assumptions.
The third problem mentioned above arises in DAI situations where the designer of the
system cannot provide the automated agent with a negotiation strategy in advance. For
example, in the data allocation scenario, finding possible dataset allocations can be done
only after the specifications of the datasets are known to the agents and thus cannot

See [521 for a detailed review of the bargaining game of Alternative Offers.
8 There are two reasons for this. Fit, there is the cost of communication and computation time spent on the
negoti~ion. Second, there is the loss of unused info~ation: until an agreement is reached, new documents
cannot be used. Thus, the servers wish to reach an agreement as soon as possible, since they receive payment
for answering queries.

S. KraudArtificial

Intelligence 94 (1997) 79-97

8.5

be supplied in advance by the designers. Construction of the strategies which are in

~uilib~urn can rely on theorems proven in advance, but can be done only when the set
of possible agreements can be defined. For such situations, there is a need to develop
low complexity computational techniques for searching for appropriate strategies by the
automated negotiators. The issue of the complexity involved in finding strategies is not
discussed in the game theory literature.
Another issue that is rarely discussed in game theory is the source of a utility function
or a set of preferences that is needed for any decision-making. In game theory, one
aspect of a definition of a game is the players utility functions or preferences, and
it is assumed that each player knows its utility function (and has some knowledge of
the utility function of its opponents). A designer of an automated agent is required to
provide
the agents with a utility function or a preference relation. Without doing so, the
game-theoretic techniques cannot be used for automats agents. In the data allocation
scenario, we have developed a complex utility function which takes into consideration
factors such as storage costs, retrieval costs, distances between servers, etc. Only then
can we apply game-theoretic techniques.
More details on our work on the strategic model of negotiation and the definition
of utility functions can be found in [39,44,45,65]. In the process of developing and
specifying the strategic model of negotiation, we have examined bilateral negotiations, as
well as multi-agent environments (more than two agents), single encounters and multiple
encounters, situations characterized both by complete and incomplete information, and
the differing impact of time on the payoffs of the participants [ 39,451. Recently, we
have also considered problems where there are two attributes to the agr~ments 1651.
While some combinations of these factors can result in minor delays in reaching an
agreement, the model nevertheless reveals an important capacity for reaching agreement
in early periods of the negotiation.9

By creating coalitions that allow them to share resources and cooperate on task
execution, autonomous agents may be able to increase their benefits. Cooperative gametheoretic models can be used to do this for self-motivated agents, each of which has
tasks it must fulfill and resources it needs to complete these tasks. though
the agents
can act and reach goals by themselves, it may be advantageous to join together.
For example, taxi drivers may own different types of cabs and therefore may have
different costs, different transportation capabilities, and different resulting payoffs. Each
taxi driver would like to increase his own benefits, but it may be in the drivers interest
to cooperate and form coalitions in order to achieve greater and more complex transportation capabilities. Game-theoretic coalition fo~ation theories can be used in the
development of automated agents that represent these drivers as they form coalitions.
Game theory [ 11,28,34,5 1,571 provides a good framework with concepts of a coalition and coalitional value and different notions of stability, but to use it, we have had
9 In 1441, it was shown how the strategic model can be used in appfications such as a hostage crisis
simulation.

86

S. Kraus/Artificial Intelligence 94 (1997) 79-97

to address three tasks: the development of explicit protocols for interaction among the
agents; the development of algorithms for coalition formation; while simultaneously
taking into account communication
costs and limited computation time. Most of the
work in game theory does not treat these issues, but only predicts how the players will
distribute the benefits, given a coalition configuration.
In [68,72] we addressed the three tasks mentioned above and presented algorithms
for coalition formation and payoff distribution in general environments. We focused on
a low complexity Kernel-oriented
[ 121 coalition formation algorithm. The properties
of this algorithm were examined via simulations. These have shown that the model
increases the benefits of the agents within a reasonable time period, and more coalition
formations provide more benefits to the agents.

3. Applying classical mechanics

to large scale agent systems

There are situations where cooperation among a large number of agents (hundred or
more) is needed. For example, the World Wide Web (WWW) consists of millions of
users and is still growing. Another example is the employment of hundreds of simple,
inexpensive autonomous mobile devices to achieve military and civilian goals in ground,
air, and underwater environments
[ 221. In such situations [ 3 1,731, the agents work
together toward satisfying a large set of joint goals, and the designers of the agents can
agree in advance on regulations and protocols for the agents interaction.
The negotiation and coalition formation methods presented in the previous section are
suitable for environments
with a relatively small number of agents. But, in very large
agent-communities,
these negotiation methods are typically too computationally
complex and time-consuming.
Furthermore, with hundreds of agents, direct communication
connections between all of the agents may be impossible or too costly to establish.
Physical models of particle-dynamics
have proved useful in such settings. They use
mathematical formulation either to describe or to predict the properties and evolution
of different states of matter. In particular, we developed efficient techniques for cooperation among hundreds of agents by adopting methods of classical mechanics used by
physicists to tackle the problem of finding the properties of interaction among many
particles. Although there are many differences between particles and computational
systems, we have shown that the classical mechanics approach yields a model that
enables feasible cooperation in very large agent-systems; the approach has a low computational complexity, which is crucial for the functioning of such systems. We have
applied the classical mechanics-based
methods to the following freight transportation
example [ 16,63,75].
Example 2 (Freight transportation system).
The system of freight transportation consists of many carriers (e.g., messengers on motorcycles)
which belong to the same
company, operating in a big city. Each carrier has a freight carrying capability that is
given in units of volume and has a given location. The tasks that the carriers must fulfill
are freight transportation tasks. We deal here with freight (e.g., packages) that should be
moved from various locations to other locations. There are many freight transportation

S. KraudArtijicial

Intelligence 94 (1997) 79-97

87

tasks to perform, and the carriers would like to perform them as soon as possible, while
at the same time minimi~ng the companys expenses.
In the above example and in the other DA1 environments that we consider, there
is a large set of agents and a large set of goals they need to satisfy. Each agent has
capabilities and should move toward satisfying goals. The first step in applying the
classical lnechani~s model to DA1 is the match between particles and their properties,
agents and their capabilities, and goals and their properties. The next step is to identify
the state of matter for modeling a community of agents and goals. The mathematical
formulation that is used by physicists either to describe or to predict the properties and
evolution of particles in these states of matter serve as the basis for the development
of algorithms for the agents. However, several modifications of the classical mechanics
model are necessary to provide an efficient algorithm for automated agents.
In the physical world, mutual attraction between particles causes motion. The reaction of a particle to the field of potential will yield a change in its coordinates and
energies. The change in the state of the particle is a result of the influence of the potential. For DAI, the agents calculate the attraction and move according to the results
of these calculations. That means, in our model, that each agent calculates the effect
of the potential field on itself by solving a set of differential equations. According
to the results of these calculations, it moves to a new state in the goal-domain. If it
reaches a goal, it will proceed to a goal-satisfaction process. In cases where too many
agents fit the requirements of the same goal, some are prevented from reaching the
goal, through the property of mutual rejection between dynamic particles. We model
the goal-satisfaction process by a collision of dynamic particles with static particles.
Because the properties of particle collisions are different from the properties of goalsatisfaction, several adjustments were made to develop efficient algorithms for agent
systems.
For example, in the freight ~~s~~ation
system of Example 2 each piece of freight
is modeled by a static particle and each carrier is modeled by a dynamic particle,
since carriers move toward the tasks location. The volume of carriers freight carrying
capabilities and the volume of each piece of freight are modeled by particle masses, and
their locations by particle locations.
The interaction between a carrier and a piece of freight is modeled by the mutual
potential function of the modeling particles. It is calculated with respect to the distance
between them. The potential functions derivatives yield forces which act on a dynamic
particle and direct it. That is, the advancement towards a piece of freight is modeled
by the movement of a dynamic particle towards a static particle. Repulsion between
two dynamic particles which model two different carriers will influence the freighttask distribution among the carriers and will prevent two carriers from proceeding to
a piece of freight which can be moved by one carrier. The performance of a freighttransportation task is modeled by the collision between a static particle, which models
the task, and a dynamic particle, which models the agent.
In [70], we provide a detailed algorithm to be used by a single agent within the
system. The algorithm leads to agent-goal allocation, and it converges to a solution where
the fulfillment of goals is accomplished either by single agents or by groups of agents

88

S, Kraus/Art$cial

Intelligence 94 (1997) 79-97

via cooperation. The computational complexity is low, and no explicit communication

is
necessary. In addition to these properties, we have proven that the algorithm we provide
performs relatively close to the optimum.
The physics approach has several advantages. While common DA1 algorithms must
be checked for their validity either by a formal proof or by simulations, the models
that are based on physics techniques can rely on theoretical and experimental results
that are already known from physics. According to these results, one can predict the
evolution of the modeled agent-system, since it will evolve in the same manner as a
corresponding
physical system. The local interactions, which enable one to derive the
global behavior of the system, assure a low computational complexity of the model. In
very large-scale agent-systems, this approach provides a model that promises emergent
cooperative goal-satisfaction
activity. In addition, the properties of the system as a whole
can be analyzed, using concepts from statistical mechanics. The employment of such
concepts enables us to derive the properties of a system through the properties of its
components.

4. Applying operations

research techniques

Many DA1 researchers have considered situations of cooperative automated agents:

for example, several workstations working together on fulfilling tasks [ 481, multi-agent
for integration of design, manufacturing and shop floor control activities [ 41, and cooperative shipping companies [ 181. In such situations, all the agents work together toward
the satisfaction of a joint goal; the designer of the automated agents can develop, in
advance, protocols for cooperation between the agents; the number of agents is not large;
and the agents can communicate and have computation capabilities.
We recommend, in such situations, the consideration of operations research techniques.
Researchers in operations research seek to determine how best to design and operate an
organizational system, usually under conditions of scarce resources [ 761.
Autonomous
agents working in DPS environments
can be considered as an organizational system, and thus algorithms that were developed for human organizations
in operations research may be applied to DAI environments.
This is suitable for environments with a few dozen agents with large computation capabilities, because the
computational complexity of the operations research techniques is usually high, and their
efficiency decreases with the size of the organization to which they are applied.
We have applied operations research techniques which were developed for the set
covering and set partitioning problems for coalition formation in DPS environments
[ 69,711. Given a set of agents and a set of tasks which they have to satisfy, we
consider situations where each task should be attached to a group of agents which will
perform the task. An example is a transportation company, similar to the example in the
previous section. The company supplies transportation services via a number of trucks,
lift trucks, cranes, boats, and planes. The drivers belong to a cooperative and share the
benefits equally, and thus try to maximize the overall benefits of the company. There may
be occasions in which one vehicle cannot perform a given transportation task by itself.
In such cases, cooperation is necessary. Therefore, several drivers will form groups, and

S. Kraus/Artificial Intelligence 94 (1997) 79-97

89

each gr0u.p will fulfill a transportation task cooperatively. If the transportation company
has many drivers, a dist~but~ task allocation m~hanism may be advantageous.
As we mentioned above, task allocation among agents may be approached as a
problem of assigning groups of agents to tasks, and, therefore, the partition of the
agents into subgroups becomes the main issue, and our problem becomes similar to
the Set Partitioning Problem (SPP). Set partitioning entails the partition of a set into
subsets, and the set petitioning problem is finding such a partition that has a minimal
cost. lo The SPP has been dealt with widely in the context of NP-hard problems 1231,
and apptoximation algorithms were developed in operations research [ 2,3,9,10,24].
Among them we can find the algorithm of Chvatal [lo], which has a logarithmic ratio
bound.
The details of the algorithm that we developed, which is based on the operations
research methods for the SPP is specified in [69]. Although the general task allocation
problem is computationally exponential, the algorithm above is polynomial and yields
results which are close to the optimal results and bounded by a logarithmic ratio bound.
Another advantage of the algorithm, which is crucial in the case of a distributed system,
is the distribution of the algorithm. We distribute the calculations in a natural way.
That is, the dis~bution is an outcome of the algorithm ch~acte~stics, since each agent
performs mostly those calculations that are required for its own actions during the
process. In addition, our distribution method prevents most of the possibly overlapping
calculations, thus saving unnecessary computational operations.
The algorithm is an anytime algorithm. If halted before normal termination, it still
provides the system with severai coalitions that have already formed. Since the first
coalitions to be formed are the better ones, the results, when halted, are still of good
quality. The anytime property of such an algorithm is important for dynamic environments, wherein the time-period for negotiation and coalition formation processes may
be changed during the process.
In another paper [SS], we considered the problem of dis~ibut~ dynamic task allocation by 3 set of cooperative agents. We modeled the agents, using a stochastic closed
queueing network, which is a well known operations research technique.
In both cases, we have developed polynomial algorithms that provide near optimal
results. From our experience, we realized that in order to apply operations research
techniques to DA& there are several steps that must be taken. First, there is the need
to find a problem that was considered in operations research which is close to the
DAI problem and to make a detailed match between the problems. For example, in the
coalition formation problem described above, we realized that it is close to the SPP
or SCP problems. Then, there is the need to adjust the operations research algorithm
to the DAI environment. In particular, most of the operations research algorithms are
centralized, and, since we deal with autonomous agents, we seek distribute algorithms.
In addition, there is the need to develop utility functions that can be used by the
agents. In operations research it is assumed that cost function is provided as part of
lo Coalition formation where coalitions may overlap cau be approached as a Set Covering Problem (SCP) _
Ii An approximation algorithm for a problem has a ratio bound p(n) if p(n) is smaller than the ratio between
the optimal cost and the approximated cost.

90

S. Kraus/Artifcial

Intelligence 94 (1997) 79-97

the problem (as in game theory). In our model, we need to provide the agents with
efficient techniques to calculate them (see also [ 641) . For example, in [ 69,7 1] we had
to develop the cost function and coalitional values in the context of task allocation and
to provide a distributed algorithm to compute them. This notion of coalitional value
is different from the notion of game-theoretic coalitional value, since here the value
depends on the coalitional configuration and on the task allocation.
Although adjusting the operations research techniques to DA1 situations required some
effort, we determined that the benefits from using these well-developed methods, and
techniques for evaluating them, may help in reaching efficient algorithms for the DA1
environment.

5. The application
automated agents

of informal

models of behavioral

and social sciences to

There are situations where automated agents need to interact with other agents in
non-structured
environments;
for example, an information server which works to form
a multi-media
document for answering a complex query of a user, agents that help
train people in negotiation [ 441, and agents that sell goods on the World Wide Web
[ 81. In such situations, the agents are self-motivated, and usually the automated agents
need to interact with people. The number of agents in the environment is not large, and
communication
is possible.
In such situations, we found that formalizing and implementing
informal models of
behavioral and social sciences can be beneficial. Behavioral and social sciences study human cooperation and coordination and develop frameworks and models of organizations
and communities
(e.g., [ 20,46,59,60]
) . In non-structured
and unpredictable environments, heuristics for cooperation and coordination among automated agents, based on
successful human cooperation and interaction techniques, may be useful.
We have applied informal models to different types of environments,
and we will
discuss one of them below. Applying informal models to DA1 can be done in two ways:
(a) using the informal models as motivation for the development of heuristics for
the cooperative activities of the automated agents;
(b) formalizing the informal models (e.g., using logic) and then applying them to a
DA1 environment.
In both cases, there is a need to carry out simulations in order to evaluate the performance
of the techniques, since the informal models usually do not formally analyze the behavior
of the systems. The main advantage in using these models is that we build upon
experience and expertise that were developed over the years in the specific type of
interactions, rather than starting from scratch and using only our own experience. Our
success in the developments of specific applications, in particular automated negotiators
[ 41,421, supports this claim.
There are two main approaches in the social sciences to the development of theorems
relating to negotiation. The first approach, which we used in Section 2.1, is the formal
theory of bargaining. This formal game-theoretic approach provides clear analyses of
various situations and precise results concerning the strategy a negotiator should choose.

S. ~raus/~~~cia~

lnre~f~gence 94 (1997) 79-97

91

However, it requires making restrictive assumptions, and the agents need to follow strict
negotiation protocols which are not possible in some real world environments.
The second approach, which we refer to as the negotiation guides approach, comprises
informal theories which attempt to identify possible strategies for a negotiator and to
assist him in achieving good results (see, for example, [ 13,19,29,33,35]).
These
negotiatmn guides do not accept the strong restrictions and assumptions presented in the
game-theoretic models. Applying these methods to DA1 is more difficult than using the
first approach, since there is no formal theory nor strategies that can be used. However,
these methods can be used in domains where people interact with each other and
with automated agents, and situations where automated agents interact in environments
without pre defined regulations. These informal models can serve as guides for the
development of negotiation heuristics [41] or as a basis for the development of a
logical model of negotiation [ 421.
In [ 37,411, we developed a general structure for a self-motivated negotiating Automated Agent acting in environments where cooperation between the agents may be
beneficial., but where conflicts among the agents can arise. There are no strict regulations and protocols for the negotiation, there is no mediator, and central controllers do
not exist. Thus agreements are not enforced, and agents may break their promises, The
agents have incomplete information concerning the other agents goals and tasks, and
an agent can provide the other agents with false information.
As a testbed, a specific domain was chosen, the ~i~lorn~cy game, which is rich
enough to include most aspects of negotiation. I2 Given a (restricted version of) natural
language which covers this domain, our agent, Diplomat, was confronted with human
agents and even demonstrated an advantage over its human negotiation partners.
The framework of Diplomat consists of five modules: the Prime Minister, that directs
the Diplomats activities; the Ministry of Defense, that is responsible for the planning;
the Foreign O&e, that negotiates with the other players; the Headquarters that executes the basic tasks of Diplomat; and the ~~teZZigenceAgency, that is responsible for
collecting information about the environment and the other players. These modules are
implemented by a dynamic set of local agents that work together, communicate, and
exchange messages to achieve the common general tasks of Diplomat.
In the design of Diplomat and in choosing the negotiation heuristics it uses, we used
different general informal negotiation guides. For example, as we mentioned above,
Diplomat consists of different modules for planning-i.e., the Ministry of Defenseand negotiations-i.e.,
the Foreign Office. The development of different modules for
negotiation and planning is a characteristic of a good negotiator, according to Fisher and
Urys mod.el [ 191. They suggest that a good negotiator should do much inventing, that
is, find out new ideas that are not already among the negotiation issues. The separation
of the planning and negotiation into two modules enables the Ministry of Defense to
find as many solutions to the problem as possible, without taking into account whether
or not they are acceptable to the other side. The ideas will not be conveyed to the other
I2 Diplomacy is a board game marketed by Avaton Hill Company and played on the map of Europe during the
years just prior to World War I. Coalitions and agreements among the players significantly affect the course
of the game.

92

S. Kraus/Artijicial Intelligence 94 (1997) 79-97

side until the Foreign Office decides to do so. Therefore, their consideration
by the
Ministry of Defense can do no harm.
There are several heuristics that Diplomat uses to decide how to make suggestions
to another agent. For example, when considering a cooperation agreement with another
agent, Diplomat designs several possible strategies and compares them to choose the
strategy that will be a basis for the agreement. Since a negotiator wants to win, one
may suspect that the only criterion that will guide him while comparing and choosing
between strategies will be his own benefits derived from the strategies. However, as
has been suggested by the literature on human negotiation, this is not the case. The
reason for that phenomena is that in order for the agreement to last, it should be
beneficial to all parties involved. Otherwise, a neglected partner may be tempted to
reach a more appealing agreement, even without informing the negotiator. For that
same reason, the other partner should be convinced that the agreement is profitable to
Diplomat (see [ 191); otherwise he will suspect that the negotiator will later break the
agreement.
In order to test Diplomat, we arranged several Diplomacy games, and our findings
(see [41] ) show that Diplomat played well in the games in which it participated.
We believe that its success is due to the integration of the heuristic techniques we
developed for the construction of negotiator agents and well developed informal theories
of negotiation, t3

6. Conclusions
In this paper we argue that applying multi-entity techniques, such as game theory and
physics, to DAI, is beneficial. We described several attempts to apply methodologies
from diverse fields to DA1 problems. A summary of the multi-entity techniques that we
used and their application in DAI is given in Table 1. The last column uses the parameters presented in the introduction to characterize the problems that we considered.
For example, we applied game theory in environments where the agents are automated
and self-motivated, but it is possible that the agents will follow some agreed-upon protocols (Sections 2.1 and 2.2). We demonstrated that classical mechanics models are
useful for task distribution in very large sets of cooperative agents (Section 3). We
applied operations research techniques such as queueing networks for task distribution
among a relatively small set of cooperative agents (Section 4). We used the less formal
social science models of cooperation when there were no strict protocols for the cooperation (Section 5), or when communication
was not possible [ 15,431. Further, we
demonstrated that ideas drawn from philosophy can be the basis for the development of
SharedPlans among agents [ 26,271.
I3 We have applied other informal models to DA1 situations. In [42], we developed a formal logic that
forms a basis for the development of a formal axiomatization
system and the implementation
of a logicbased negotiator [ 141 based on persuasion models [ I]. In [ 26,271, we have applied philosophical informal
models of cooperative activity [7] for situations where teams composed of people and computers plan and
work together toward satisfying a shared goal. In [ 15,431, we used the notion of focal point introduced by
&helling [ 58,661, for multi-agent cooperation without communication.

S. Kraus/Art$cial

Intelligence

94 (1997) 79-97

93

Table 1
Summary of multi-entity techniques and their application in DAL In the last column, SMA stands for selfmotivated agents, and CA indicates cooperative agents which work toward satisfying the same goal (see
Section 1) . R&P indicates that the designers can agree on regulations and protocols for agents interaction. s#,
m# and I# srands for environments with small (handful), medium (few dozen), or large number (hundreds)
of agents, respectively. AUTO indicates environments with only automated agents, and AUTO&PE stands for
systems composed of people and automated agents. That communication is possible is indicated by COMU.
Multi-en&v techniaues

DA1

Paoers

Characterization

Negotiation for task distribution

& Resource allocation in MA

144,451
L391

SMA, s#, R&P,

AUTO, COMU

Coalition formation in MA

168,721

SMA, m#, R&P,

AUTO, COMU

Con~cting tasks in MA

138,401

SMA, s#, R&P,

AUTO, COMU

Goal satisfaction in very

Iarge DPS environments

[701

CA, I#, R&P,

AUTO

Coalition formation in DPS

169,711

CA, m#, R&P,

AUTO, COMU

Queueing networks

Task allocation in DPS

[551

CA, m#, R&P,

AUTO, COMU

Behavioral sciences
Negotiation guides

Diplomatic negotiation

137,411

SMA, m#,
AUTO&PE, COMU

Persuasion models (logic)

Argumentation

[ 14,421

SMA, s#,
AUTO&P& COMU

Focal points

Cooperation without
communication

[ 15,431

CA, m#, AUTO,

R&P

Collaborative plans

126,273

CA&WA, m#,
AUTO&PE, COMU

Game theory
Strategic bargaining models
Theories of coalition
formation
~inciple-~lgent models

Physics
Classical mechanics

Operations research
SPP & SCP

(logic & decision theory)

Philosophy (logic)

There are two main aspects of a multi-entity environment that determine its usefulness
to a DA1 problem and its effect on the amount of work required for the adaptation of
techniques developed for it to the DAI problems. The first criterion is the similarity
between the entities and the automated agents. The second criterion is the level of
fo~aIizati(~n
that is used by researchers of the multi-entity domains.
For example, people are more similar to automats agents than are particles. Therefore,
in all the multi-entity
techniques that were developed for humans environments,
it
was not difficult to match the entities in the environment and the participants in the
multi-agent domains. For example, it is clear that players in game-theoretic frameworks
can model automated agents. It is Iess clear which types of particles in the cIassica1
mechanics framework serve as models for agents and that collisions are a good way to
model goal-satisfaction.

94

S. Kraus/Artificial Intelligence 94 (I 997) 79-97

The second criterion has to do with the fact that we need to provide our automated
agents with formal and well-designed algorithms. With respect to this, it is easier to use
techniques from formal multi-entity
models than techniques that were not formalized
by their developers. For example, even though people and automated agents have much
in common, with respect to cooperation, it is quite difficult to develop an algorithm for
agent cooperation based on the informal ideas, procedures, and rules that are presented by
social scientists and philosophers. Much effort is required to formalize these procedures
and rules and to produce an implementable
algorithm for the automated agents. On
the other hand, after going through the process of modeling a community of agents
using a classical mechanics framework, the usage of the formal techniques of classical
mechanics is not so difficult. There is a need to modify the formal procedures and to
adjust them to the multi-agent requirement, but there is no need to create the formal
procedure from scratch.

Acknowledgments
I would like to thank the many people who, over the years, have collaborated with
me: C. Bat-al, E. Blake, P. Bonatti, E. Ephrati, A. Evenchik, D. Etherington, M. Fenster,
B. Grosz, M. Harris, J. Hendler, K. Holley, J. Horty, D. Lehmann, G. Lemel, M. Magidor,
J. Minker, M. Nirkh, D. Perlis, T. Plotkin, J. Rosenschein, A. Schwartz, 0. Shehory,
Y. Shoham, S. Subrahmanian,
K. Sycara, B. Thomas, J. Wilkenfeld, and G. Zlotkin. Our
joint work influenced my thinking on cooperation and coordination.
I would like to thank Barbara Grosz, Martha Pollack, Jonathan Wilkenfeld, Onn
Shehory and Orna Schechter, each of whom also provided help and support while I
was preparing the Computers and Thought lecture and this paper. Special thanks to Dr.
Shifra Hochberg for editorial assistance.
This work was supported by the NSF under Grants No. IRI-9423967 and IRI-9311988
and the Israeli Ministry of Science, Grants No. 6288 and 4210.

References
[ I ] H. Abelson, Persuasion (Springer, New York, 1959).
[2] E. Balas and M. Padberg, On the set covering problem, Oper. Res. 20 (1972) 1152-l 161.
[ 31 E. Balas and M. Padberg, On the set covering problem: an algorithm for set partitioning, Oper. Res. 23
(1975) 74-90.
[4] S. Balasubramanian
and D. Norrie, A multi-agent intelligent design system integrating manufacturing
and shop-floor control, in: Proceedings 1st International Conference on Multiagent Systems ( 1995)
3-19.
151 T. Balch and R.C. Arkin, Motor schema-based
formation control for multiagent robot teams, in:
Proceedings 1st International Conference on Multiagent Systems (1995) 10-16.
[6] A.H. Bond and L. Gasser, An analysis of problems and research in DAI, in: A.H. Bond and L. Gasser,
eds., Readings in Distributed Artificial Intelligence (Morgan Kaufmann, San Mateo, CA, 1988) 3-35.
[71 M.E. Bratman, Shared cooperative activity, Philosophical Review 101 (1992) 327-341.
[ 81 A. Chavez and P Maes, Kasbah: an agent marketplace for buying and selling goods, in: Proceedings 1st
International Conference on the Practical Application of IntelligentAgents and Multi Agents Technology,
London (1996)75-90.

S. Kraus/Artijicial Intelligence 94 (1997) 79-97

95

[9] N. Christofides and S. Korman, A computational survey of methods for the set covering problem, Math.
Oper. Res. 21 (1975) 591-599.
[ lo] V. Chvatal, A greedy heuristic for the set-covering problem, Math. Oper. Res. 4 ( 1979) 233-235.
[ 1 l] M.S.Y. Chwe, Farsighted coalitional stability, J Economic Theory 63 (1994) 299-325.
[ 121 M. Davis and M. Maschler, The kernel of a cooperative game, Naval Res. Logist. Quart. 12 (1965)
223-2.59.
[ 131 D. Dnrckman, Negotiations (Sage, Beverly Hills, CA, 1977).
[ 141 A. Evenchik, Inference system for argumentation in negotiation between automatic agents, M.Sc. Thesis,
Department of Mathematics and Computer Science, Bar-&n University, Ramat Gan, Israel ( 1995).
[ 151 M. Fenster, S. Kraus and J. Rosenschein, Coordination without communication: experimental validation
of focal point techniques, in: Proceedings Jst International Conference on Multiagent Systems ( 1995)
102-l 16.
[ 161 K. Fischer and N. Kuhn, A DA1 approach to modeling the transportation domain, Technical Report RR
93-25, Deutsches Forschungszentrum
fiir Kiinstliche Intelligenz GMBh (1993).
[ 171 K. Fischer, J.P. Milller, I. Heimig and A. Scheer, Intelligent agents in virtual enterprises, in: Proceedings
1st In!ernational Conference on the Practical Application of Intelligent Agents and Multi Agents
Technology. London, 1996.
[ 181 K. Fischer, J.P. Milller, M. Pischel and D. Schier, A model for cooperative transportation scheduling, in:
Proceedings 1st International Conference on Multiagent Systems ( 1995) 109-l 16.
[ 191 R. Fisher and W. Ury, Getting to Yes: Negotiating Agreement without Giving in (Houghton Mifllin,
Boston, MA, 1981).
[20] R.C. Ford, R.B. Armandi and CF? Heaton, Organization Theory: An Integrative Approach (Harper and
Row, New York, 1988).
[ 211 D. Fudenberg and J. Tirole, Game Theory (MIT Press, Cambridge, MA, 1991).
[22] D.W. Gage, Command control for many-robot systems, Unmanned Systems (Fall, 1992) 28-34.
[23] M.R. Garey and D.S. Johnson, Computers and Intractability: A Guide to the Theory of NP-Completeness
(Freeman, New York, 1979.
[24] R.S. Garfinkel and G.L. Nemhouser, The set-partitioning problem: set covering with equality constraints,
Oper. .Res. 17 (1969) 848-856.
[25] L. Glicoe, R. Staats and M. Huhns, A multi-agent environment for department of defense distribution,
in: Proceedings IJCAI-95 Workshop on Intelligent Systems, Montreal, Que. ( 1995).
[26] B. Grclsz and S. Kraus, Collaborative plans for group activities, in: Proceedings IJCAI-93, Chambery,
France (1993) 367-373.
[27] B.J. Grosz and S. Kraus, Collaborative plans for complex group activities, Artificial Intelligence 86
(1996, 269-357.
[ 281 S. Gui;lsu and M. Malitza, Coalition and Connection in Games (Pergamon, Oxford, 1980).
[ 291 Lavinia Hall, ed., Negotiation: Strategies for Mutual Gain (Sage, Beverly Hills, CA, 1993).
1301 J.C. Harsanyi,
Rational Behavior and Bargaining Equilibrium in Games and Social Situations
(Cambridge University Press, Cambridge, 1977).
[ 311 T. Hogg, Social dilemmas in computational
ecosystems, in: C.S. Mellish, ed., Proceedings IJCAI-95,
Montreal, Que. (Morgan Kaufmann, San Mateo, CA, 1995) 711-718.
[32]
[ 331
[ 341
[35]

[ 361
[37]
[38]

N.R. Jennings, Controlling cooperative problem solving in industrial multi-agent systems using joint
intentions, Artificial Intelligence 75 (1995) l-46.
R. Johnson, Negotiation Basics (Sage, Beverly Hills, CA, 1993).
J.P. Kahan and A. Rapoport, Theories of Coalition Formation (Lawrence Erlbaum, Hillsdale, NJ, 1984).
C.L K..urass, The Negotiating Game: How to Get What You Want (Thomas Crowell Company, New
York, 1970).
M. Klusch and 0. Shehory, A polynomial kernel-oriented
coalition formation algorithm for rational
information agents, in: Proceedings ICMAS-96, Kyoto, Japan ( 1996).
S. Kraus, Planning and communication
in a multi-agent environment, Ph.D. Thesis, Hebrew University,
Jerusalem, 1988 (written largely in Hebrew).
S. Kraus, Agents contracting
tasks in non-collaborative
environments,
in: Proceedings AAAI-93,
Washington, DC ( 1993) 243-248.

S. Kraus/Arttjicial Intelligence 94 (1997) 79-97

96

[391 S. Kraus, Beliefs, time and incomplete information in multiple encounter negotiations

among autonomous
agents, Ann. Math. Artif: Intell. ( 1997).
1401 S. Kraus, An overview of incentive contracting, Artificial Intelligence 83 (1996) 297-346.
1411 S. Kraus and D. Lehmann, Designing and building a negotiating automated agent, Comput. Intell. 11
(1995) 132-171.
1421 S. Kraus, N. Nirkhe and K.P Sycara, Reaching agreements through argumentation: a logical model, in:
Proceedings DAI-93 ( 1993).

I431 S. Kraus and J.S. Rosenschein,

[441
[451
[461
[471
[48]

[49]
[50]

The role of representation in interaction: discovering focal points among

alternative solutions, in: E. Werner and Y. Demazeau, eds., Decentralized Art$cial Intelligence, Vol. 3
(Elsevier, Amsterdam, 1992) 147-165.
S. Kraus and J. Wilkenfeld, A strategic negotiations model with applications to an international crisis,
IEEE Trans. Systems Man Cybernet. 23 (1993) 313-323.
S. Kraus, J. Wilkenfeld and Cl. Zlotkin, Multiagent negotiation under time constraints, Artificial
Intelligence 75 (1995) 297-345.
W.A. Kraus, Collaboration in Organizations: Alternatives to Hierarchy (Human Sciences Press, New
York, 1984).
R.D. Lute and H. Raiffa, Games and Decisions (John Wiley and Sons, New York, 1957).
T.W. Malone, R.E. Fikes, K.R. Grant and M.T. Howard, Enterprise: a marketlike task schedule for
distributed computing environments,
in: B.A. Huberman, ed., The Ecology of Computation (NorthHolland, Amsterdam, 1988) 177-205.
R. Myerson, Game Theory (Harvard University, 1991).
NASA, EOSDIS home page, http://www-vOims.gsfc.nasa.gov/vOims/index.html,
1996.

Theory of Games and Economic Behavior (Princeton University

Press, Princeton, NJ, 1947).
1521 M.J. Osborne and A. Rubinstein, Bargaining and Markets (Academic Press, San Diego, CA, 1990).
L531 M.J. Osborne and A. Rubinstein, A Course in Game Theory (MIT Press, Cambridge, MA, 1994).
[541 C. Petrie, M. Cutlosky and H. Park, Design space navigation as a collaborative aid, in: Proceedings 3rd
International Conference on Artificial Intelligence in Design, Lausanne ( 1994).
[551 T. Plotkin and S. Kraus, A queueing network approach to distributed task allocation for cooperative
agents (submitted, 1996).
[561 M. Pollack, T. Znati, E. Ephrati, D. Joslin, S. Lauzac, A. Nunes, N. Onder, Y. Ronen and S. Ur, The
dipart project, in: Proceedings ARPI Annual Meeting, Tucson, AZ ( 1994).
[571 A. Rapoport, N-Person Game Theory (University of Michigan, Ann Arbor, MI, 1970).
1581 E. Rasmusen, Games and Information (Basil Blackwell Ltd., Cambridge, MA, 1989).
1591 M.1. Reed, The Sociology of Organisations: Themes, Perspectives and Prospects (Harvester Wheatsheaf,
New York, 1992).
[601 S.P Robbins, Organization Theory (Prentice Hall, Englewood Cliffs, NJ, 1990).

1511 J. von Neumann and 0. Morgenstem,

and G. Zlotkin, Rules of Encounter: Designing Conventions for Automated Negotiation

Among Computers (MIT Press, Boston, MA, 1994).
A. Rubinstein, Perfect equilibrium in a bargaining model, Economefrica 50 (1982) 97-109.
T. Sandholm, An implementation
of the contract net protocol based on marginal cost calculations, in:
Proceedings AAAI-93, Washington, DC (1993) 256-262.
T.W. Sandholm and V.R. Lesser, Coalitions among computationally bounded agents, Artificial Intelligence
94 (1997) 99-137 (this issue).
0. Schechter, Sharing resources through negotiation in multi-agent environments, Masters Thesis, BarBan University, Ramat-Gan, Israel ( 1996).
T.C. Schelling, The Strategy of Conflict (Oxford University Press, New York, 1963).
S. Sen and E.H. Durfee, A formal study of distributed meeting scheduling, in: Group Decision and
Negotiation Support Systems ( 1997).
0. Shehory and S. Kraus, Coalition formation among autonomous agents: strategies and complexity, in:
From Reactions to Cognition (Springer, Berlin, 1995) 57-72.
0. Shehory and S. Kraus, Task allocation via coalition formation among autonomous agents, in:
Proceedings IJCAI-95, Montreal, Que. (1995) 655-661.

[611 J.S. Rosenschein

[621
I631
[641
[651
[661
[671
[681
[69]

S. Kraus/Arti&ial

Intelligence 94 (1997) 79-97

97

[70] 0. Shehory and S. Kraus, Emergent cooperative goal-satisfaction

in large scale automated-agent
systems,
in: Proceedings ECAI-96, Budapest ( 1996) 544-548.
[ 711 0. Shehory and S. Kraus, Formation of overlapping coalitions for precedence-ordered task-execution
among autonomous agents, in: Proceedings ICMAS-96 ( 1996) 330-337.
[72] 0. Shehory and S. Kraus, A kernel-oriented
model for coalition-formation
in general environments:
Implementation and results, in: Proceedings AAAI-96, Portland, OR (1996) 134-140.
[73] Y. Shoham and M. Tennenholtz, On the emergence of social conventions: modeling, analysis and
simulations, Artificial Intelligence 94 ( 1997) 139-166 (this issue).
[74] M.B. Tsvetovatyy
and M. Gini, Toward a virtual marketplace:
architectures
and strategies,
in:
Proceedings 1st International Conference on the Practical Application of Intelligent Agents and Multi
Agem Technology. London (1996) 597-614.
[75] MI? Wellman, A market-oriented
programming
environment
and its application
to distributed
multicommodity
flow problems, J. Artif: Intell. Res. 1 (1993) l-23.
[ 761 W.L. Winston, Operations Research: Applications and Algorithms ( PWS-Kent Publishing Company,
Boston, MA, 1987).