International Journal of Intelligence Science, 2014, 4, 1-6
Published Online January 2014 (
A Cognitive Model for Multi-Agen t C ollabor ation
Zhongzhi Shi1, Jianhua Zhang1,2, Jinpeng Yue1,2, Xi Yang1
1Key Laboratory of Intelligent Information Processing, Institute of Computing Technology,
Chinese Academy of Sciences, Beijing, China
2University of Chinese Academy of Sciences, Beijing, China
Received October 27, 2013; revised November 23, 2013; accepted November 30, 2013
Copyright © 2014 Zhongzhi Shi et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. In accor-
dance of the Creative Commons Attribution License all Copyrights © 2014 are reserved for SCIRP and the owner of the intellectual
property Zhongzhi Shi et al. All Copyright © 2014 are guarded by low and by SCIRP as a guardian.
In multi-agent system, agents work together for solving compl ex tasks and reach i ng comm on g oals. In thi s paper,
we propose a cognitive model fo r multi-agent collaboration. Based on the cognitive model, an agent architecture
will also be presented. This agent has BDI, awareness and policy driven mechanism concurrently. These ap-
proaches are integrated in one agent that will make multi-agent collaboration more practical in the real world.
Cognitive Model; Multi -Agent Collaboration ; Aw are ness; ABGP Model; Policy Driven S trategy
1. Introduction
Collaboration occurs over time as organizations interact
formally and informally through repetitive sequences of
negotiation, development of commitments, and execution
of those commitments. Both cooperation and coordination
may occur as part of the early process of collaboration,
and collaboration represents a longer-term integrated pro-
cess. Gray describes collaboration as “a process through
which parties who see different aspects of a problem can
constructively explore their differences and search for
solutions that go beyond their own limited vision of what
is possible[1].
Agent is an active entity and chooses a decision
process dynamically through interactive with environ-
ment [2]. The agent makes a choice from a large set of
possibilities which can determine the transformation of
environmental information, the internal information and
decision process. Thus the result is the mode of operation
that an agent choose. In multi-agent system, the agent
performs one or multiple tasks. In other words, the inter-
nal information from an agent is also a function of the
information that it receives from other agents. This
represents the problem in a new way i.e. the multi-agent
takes decision collaboratively for which it optimizes the
team behavior. Thus individual decision is avoided for
the global interest irrespective of environmental input.
Hence a better action is chosen in a collaboration manner
according to the environment information.
In a multi-agent system, no single agent owns all
knowledge required for solving complex tasks. Agents
work together to achieve common goals, which are
beyond the capabilities of individual agent. Each agent
perceives information from the environment with sensors
and finds out the number of cognitive tasks, selects the
particular combination for execution in an interval of
time, and finally outputs the required effective actions to
the environment. There is one significant prior work
formalizing joint activities. Paper [3] provides analyses
for Bratmans three characteristic functional roles pla yed
by intentions, and shows how agents can avoid intending
all the foreseen side-effects of what they actually intend.
Paper [4] describes collaborative problem-solving and
mixed initiative planning. However, these models focus
more on formal aspects of belief states and reasoning
rather than how agents behave. Other works, such as
COLLAGEN [5] and RavenClaw [6] focus on task ex-
ecution but lack explicit models of planning or commu-
nication. The PLOW system [7] defines an agent that can
learn and execute new tasks, but the PLOW agent is de-
fined in procedural terms making it difficult to generalize
to other forms of problem-solving behavior. Paper [8]
builds cognitive model on the theories, accounts for col-
laborative behavior including planning and communica-
tion, and in which tasks are represented declaratively to
support introspection and the learning of new behaviors,
but only focuses on internal mental state of agent and
does not consider environment situation.
As an internal mental model of agent, BDI model has
been well recognized in philosophical and artificial intel-
ligence area. Bratmans philosophical theory was forma-
lized by Cohen and Levesque [3]. In their formalism, in-
tentions are defined in terms of temporal sequences of
agents beliefs and goals. Rao and Georgeff have pro-
posed a possible-worlds formalism for BDI architecture
[9]. The abstract architecture they proposed comprises
three dynamic data structures representing the agents
beliefs, desires, and intentions, together with an input
queue of events. The architecture allows update and query
operations on the three data structures [10]. The update
operations on beliefs, desires, and intentions are subject to
respective compatibility requirements. These functions
are critical in enforcing the formalized constraints upon
the agents mental attitudes. The events that the system
can recognize include both external events and internal
Wooldridge and Lomuscio proposed a multi-agent
VSK logic which allows us to represent what is objec-
tively true of an environment, what is visible, or knowa-
ble about the environment to individual agents within it,
what agents perceive of their environment, and finally,
what agents actually know about their environment [11].
In multi-agent syste m. group awareness is an understand-
ing of the activities of others and provides a context for
own activity. Group awareness can be divided into basic
questions about who is collaborating, what they are doing,
and where they are working [12]. When collaborators can
easily gather information to answer these questions, they
are able to simplify their verbal communication, able to
better organize their actions and anticipate one anothers
actions, and better able to assist one another.
Policy-Based Management is a management paradigm
that separates the rules governing the behavior of a sys-
tem from its functionality. It promises to reduce main-
tenance costs of information and communication systems
while improving flexibility and runtime adaptability. A
rational agent is any entity that perceives and acts upon
its environment, and selects actions based on the infor-
mation receiving from sensors and built-in knowledge
which can maximize the agents objective. Paper [13]
points out following notions: A reflex agent uses if-then
action rules that specify exactly what to do under the
current condition. In this case, rational behavior is essen-
tially compiled in by the designer, or somehow pre-com-
puted. A goal-based agent exhibits rationality to the de-
gree to which it can effectively determine which actions
to take to achieve specified goals, allowing it greater
flexibility than a reflex agent. A utility-based agent is
rational to the extent that it chooses the actions to max-
imize its utility function, which allows a finer distinction
among the desirabilities of different states than do goals.
These three notions of agent hood can fruitfully be codi-
fied into three policy types for multi-agent system.
In this paper, a cognitive model for multi-agent colla-
boration will be pr oposed in terms of external perception
and internal mental state of agents. The agents thus are in
a never-ending cycle of perception, goal selection, plan-
ning and execution.
2. Cognitive Model ABGP
Agent can be viewed as perceiving its environment in-
formation through sensors and acting environment
through effectors [14]. A cognitive model for multi-agent
collaboration should consider external perception and
internal mental state of agents. Awareness is knowledge
created through interaction between an agent and its en-
vironment. Endsley pointed out awareness has four basic
characteristics [15]:
Awareness is knowledge about the state of a particular
Environments change over time, so awareness must be
kept up to da t e .
People maintain their awareness by in teracting with the
Awareness is usually a secondary goalthat is, the
overall goa l i s not simply
Gutwin et al. proposed a conceptual framework of
workspace awareness that structures thinking about
groupware interface support. They list elements for the
conceptual framework [16]. Workspace awareness in a
particular situation is made up of some combination of
these elements.
For internal mental state of agents we can consult BDI
model which was conceived by Bratman as a theory of
human practical reasoning. Its success is based on its
simplicity reducing the explanation framework for com-
plex human behavior to the motivational stance [17].
This means that the causes for actions are always related
to human desires ignoring other facets of human cogni-
tion such as emotions. Another strength of the BDI mod-
el is the consistent usage of folk psychological notions
that closely correspond to the way people talk about hu-
man behavior.
In terms of above considerations, we propose a cogni-
tive model for multi-agent collaboration through 4-tuple
<Awareness, Belief, Goal, Plan>. And the cognitive
model can be called ABGP model.
2.1. Awareness
Multi-agent awareness should consider basic elements
and relationship in multi-agent system. Multi-agent
awareness model is defined as 2-tuple MA = {Element,
Relation}, where Elem ents of awareness as follows:
Identity (Role): Who is participating?
Location: Where are they?
Intent i ons : What a re they going t o do?
Actions: What a re they doing?
Abilities: What can they do?
Objects: What objects are they using?
Tim e point: When do the action execute?
Basic relationships contain task relationship, role rela-
tionship, operation relationship, activity relationship and
cooperat i on r e l a tionships.
Task relationships define task decomposition and
composition relationships. Task involves activities with a
clear and unique role attribute
Role relationships describe the role relationship of
agents in the multi-agent activities.
Operation relationships describe the operation set of
Activity relationships describe activity of the role at a
Cooperation relationships describe the interactions be-
tween agents. A partnership can be investigated through
cooperation activities relevance among agents to ensure
the transmissio n of information between different p ercep-
tion of the role and tasks for maintenance of the entire
mul ti -agent perception
2.2. Belief
Belief represents the information, an agent has the world
it inhabits and its own internal state. This introduces a
personal world v iew inside the agent. Belief can often be
seen as knowledge base of an agent, which contains ab-
undant contents, including basic axioms, objective facts,
data and so on.
Belief knowledge base is a 3-triple K = <T, S, B>,
where, T describes the basic concepts in the field and
their definitions, axioms form domain concepts, namely
domain ontology; S is the domain between facts and
formulas there is a causal relationship between con-
straints, called causality constraint axiom, which ensures
consistency and integr ity of the knowledge base; B is the
set of beliefs in current state, containing facts and data.
The contents of B is changed dynamically.
2.3. Goal
Goals represent the agents wishes and drive the course of
its actions. A goal deliberation process has the task to
select a subset of consistent desires. In a goal-oriented
design, different goal types such as achieve or maintain
goals can be used to explicitly represent the states to be
achieved or maintained, and therefore the reasons, why
actions are executed. Moreover, the goal concept allows
modeling agents which are not purely reactive i.e., only
acting after the occurrence of some event. Agents that
pursue the i r o wn goals exhibit pro-act i ve behavior.
There are three ways to generate goals:
System designer sets the goal or the goal will be se-
lected when the system initializing.
Generate the goal according to observing the dynamic
of the environm e nt .
Goals generate mainly stem from internal state of
2.4. Plan
When a certain goal is selected, agent must looking for an
effective way to achieve the goal, and sometimes even
need to modify the existing goal, this reasoning process is
called planning. In order to accomplish this plan reason-
ing, agent can adopt two ways: one is using already pre-
pared plan library, which includes some of the planning
for the actual inference rules which also called static
planning; another way is instantly planning, namely dy-
namic pr ogra mming.
Static plan is aimed at some specific goal, pre- estab-
lished goals needed to achieve these basic processes and
methods, thus forming the corresponding goals and plan-
ning some of the rules, that planning regulations. Since
these rules in the system design has been written, the
planning process actually becomes a search in the library
in the pla nning proc e s s of matchi ng.
Dynamic plan finds an effective way to achieve the
goals and means for a certain goal which based on the
beliefs of current status, combined with the main areas of
the axioms and their action description. In the dynamic
situation, the environment domain is dynamically chang-
ing, the beliefs are also constantly changing. Even for the
same goal, in a different state, the planning and imple-
mentation of processes may be different. Therefore, the
dynamic plan is very important for multi-agent system,
especially i n a complex environme nt.
In dynamic mode, we use top-down methods. Typically
a goal might be composed by a number of sub-goals, so
we must fi rst achieve all sub-goals of the goal.
3. Policy-Driven Strategy
There are many definitions about the policy, and differ-
nent area has different standard. For example, IETF/
DMTF policy is defined as a series of management to a
set of rules [18]. Many people regard the policy simply
equated with defined rules, which is obviously too narrow.
We use relatively bro ad, loosely definition: p olicy is used
to guide the behavior of the system means. In the mul-
ti-agent system, policy describes agent behavior that mu st
be followed, which reflects the human judgment. Policy
tells agent what should do (ob je c tives), how to do (action),
what exte nt (utilit y) which guide the age nt behavi or.
Paper [13] defines a unified framework for autonomic
computing policies which are based upon the notions of
states and actions. In general, one can characterize a sys-
tem, or a system component, as being in a state S at a
given moment in time. Typically, the state S can be de-
scribed as a vector of attributes, each of which is either
measured directly by a sensor, or perhaps inferred or syn-
thesized from lower-level sensor measurements. A policy
will directly or indirectly cause an action “a” to be taken,
and the result of which is that the system or component
will make a deterministic or probabilistic transition to a
new state. This unif ied framework also fits to multi-agent
A multi-agent system at a time t in state s0, s0 usually
consists of a series of attributes and values to characterize.
Policy actions directly or indirectly caused by the execu-
tion of action, cau sing the system to sh ift to a new state s,
as shown in Figure 1 . Policy can be seen as a state transi-
tion function, and multi-agent system changes their state
based on poli ci es.
A policy P is defined as four-tuples P = {St, A, Sg, U},
where St is the trigger state set, th at is, the state of impl e-
mentation of the policy is triggered collection; A is action
set; Sg is the set of goal states; U is the goal state utility
function se t, to assess the merits of th e goal state level. At
least three types of policy will be useful for multi-agent
3.1. Action Policies
Action policy describes the action that should be taken
whenever the system is in a given current state. Action
policy does not take appropriate action after describing
the system achieved state, that does not give, nor given
goal state utility f unction . Typically action policy P = {St,
A, _, _}, where “_” indicates empty. Action takes the
form of IF (Condition) THEN (Action). Condition speci-
Figure 1. Multi-agent policy model based on the state tran-
fies either a specific state or a set of possible states that
all satisfy the given Condition. Note that the state that
will be reached by taking the given action is not specified
3.2. Goal Policies
Goal policy does not give the system a state should take
action, but to describe the system required to achieve the
goal state. Goal policies specify either a single desired
state, or one or more criteria that characterize an entire
set of desired states. Unlike action strategies as depen-
dent people come clear that the action taken, but goal
policy according to the goal to generate a reasonable ac-
tion. Usually goal policy P = {St, _, Sg, _}.
3.3. Utility Function Policies
A Utility Function policy is an objective function that
expresses the value of each possible state. Utility func-
tion policies generalize goal policies. Utility function
policies provide a more detailed and flexible mechanism
than the goal policies and action policies, but the utility
policies need policy makers on the system model which
has a more in-depth and detailed understanding of the
need for more modeling, optimization and algorithms.
Typically utility function policy P = {St, _, _, U}.
4. Multi-Agent Collaboration
Multi-agent systems are computational systems in which
a collection of loosely autonomous agents interact to
solve a given problem. As the given problem is usually
beyond the agents individual capabilities, agent need to
exploit its ability to collaboration, communicate with its
neighbors. For multi-agent collaboration there are two
general approaches, one approach is through awareness
facility to specify the cooperation relationship between
agents; another approach is through joint intention.
As above mentioned cooperation relationships of
awareness in ABGP model, we describe the interactions
between agents. Agent collaboration is defined explicitly
when multi-agent is designed. A partnership can be in-
vestigated through cooperation activities relevance be-
tween agents to ensure the transmission of information
between different perception of the role and tasks for
maintenance of the entire multi-agent perception.
In the joint intention theory, a team is defined as a set
of agents having a shared objective and a shared men-
tal state. Joint inten tions are held by the team as a w hole,
and require each team member to informing other one
whenever it detects the goal state change, such as goal
achieved, or it is impossible to archive or as the dynamic
of the environment, the goal is no relevant. For more
detail you can refer to paper [19]
5. Agent Architecture
In terms of the cognitive model for multi-agent collabo-
ration an agent architecture has been proposed shown in
Figure 2. The abstract architecture we propose comprises
four dynamic data structures representing the agent’s
awareness, belief, desire, and plan, together with an input
queue of events. We allow update and query operations
on the four data structures. The update operations on
awareness, beliefs, desires, and intentions are subject to
respective compatibility requirements. These functions
are critical in enforcing the formalized constraints upon
the agent’s mental attitudes, also observing the environ-
ment situation. The events of the system includes include
both external events and internal events. We assume that
the events are atomic and are recognized after they have
occurred. Similarly, the outputs of the agent actions are
also assumed to be atomic.
The main interpreter loop is given below. We assume
that the event queue, awareness, belief, desire, and inten-
tion structures are global.
options := option-generator(event-queue);
selected-options := de lib erate(options);
update-plan s (selected-options);
drop-impo ssible-attitudes();
e n d repeat
This agent architecture has three outstanding features:
Agent has BDI reasoning which is a successful ap-
proach for multi-agent systems.
Agent makes deliberation not only depending on in
Figure 2. Agent architecture.
ternal mental state, it also concerns awareness informa-
tion from outside environment. This feature makes agent
collaborative more reality.
Policy-driven reasoning can improve the performance
and enhance flexibility.
6. Conclusions and Future Works
In multi-agent system, agents work together for solving
complex tasks and reaching common goals. A cognitive
model for multi-agent collaboration is proposed in this
paper. Based on the cognitive model, we develop an
agent architecture which has BDI, awareness and policy
driven mechanism concurrently. These approaches are
integrated in one agent that will cause multi-agent colla-
boration in a more practical world.
We are going to develop a multi-agent collaboration
software system, and apply it to animal robot collabora-
tion work in the future.
This work is supported by the National Program on Key
Basic Research Project (973) (No. 2013CB329502), Na-
tional Natural Science Foundation of China (No.
61035003, 60933004, 61202212, 61072085), National
High-tech R&D Program of China (863 Program) (No.
2012AA011003), National Science and Technology Sup-
port Program (2012BA107B02).
[1] B. Gray, “Collaborating: Finding Common Ground for
Multiparty Problems,” Jossey-Bass, San Francisco, 1989.
[2] Z. Z. Shi, “Advaced Artificial Intelligence,” World Scien-
tific Publishing Co., Singapore City, 2011.
[3] P. R. Cohen and H. J. Levesque, “Intention Is Choice
with Commitment,” Artificial Intelligence, Vol. 42, No.
2-3, 1990, pp. 213-361.
[4] G. Ferguson and J. F. Allen, “Mixed-Initiative Dialogue
Systems for Collaborative Problem-Solving,” AI Maga-
zine, Vol. 28, No. 2, 2006, pp. 23-32.
[5] C. Rich and C. L. Sidner, “COLLAGEN: When Agents
Collaborate with People,” 1st International Conference
on Autonomous Agents, Marina del Rey, 5-8 February
1997, pp. 284-291.
[6] D. Bohus and A. Rudnicky, “The RavenClaw Dialog
Management Framework: Architecture and Systems,”
Computer Speech & Language, Vol. 23, No. 3, 2009, pp.
[7] J. Allen, N. Chambers, G. Ferguson, L. Galescu, H. Jung,
M. Swift and W. Taysom,PLOW: A Collaborative Task
Learning Agent,” Proceedings of the 22nd Conference on
Artificial Intelligence (AAAI-07), Vancouver, 22-216 July
2007, pp. 1514-1519.
[8] G. Ferguson and J. Allen, “A Cognitive Model for Colla-
borative Agents,” Proceedings of the AAAI 2011 Fall
Symposium on Advances in Cognitive Systems, Washing-
ton DC, 4-6 November 2011, pp. 112-120.
[9] A. S. Rao and M. P. Georgeff, “Modeling Rational
Agents within a BDI-Architecture,” In: J. Allen, R. Fikes
and E. Sandewall, Eds., Proceedings of the 2nd Interna-
tional Conference on Principles of Knowledge Represen-
tation and Reasoning, Morgan Kaufmann Publishers, San
Mateo, 1991.
[10] L. Braubach and A. Pokahr, “Developing Distributed
Systems with Active Components and Jadex,Scalable
Computing: Practice and Experience, Vol. 13, No. 2,
2012, p. 100.
[11] M. Wooldridge and A. Lomusc io , “Multi-Agent VSK Lo-
gic,” Proceedings of the 17th European Workshop on
Logics in Artificial Intelligence (JELIAI-2000), Springer-
Verlag, Berlin, 2000.
[12] C. Gutwin and S. Greenberg, “A Descriptive Framework
of Workspace Awareness for Real-Time Groupware,
JCSCW, Vol. 11, 2002, pp. 411-446.
[13] J. O. Kephart and W. E. Walsh, “An Artificial Intelli-
gence Perspective on Autonomic Computing Policies,
Proceedings of the 5th IEEE International Workshop on
Policies for Distributed Systems and Networks, Yorktown
Heights, 7-9 June 2004, pp. 3-12.
[14] Z. Z. Shi, X. F. Wang and J. P. Yue, “Cognitive Cycle in
Mind Model CAM,International Journal of Intelligence
Science, Vol. 1, No. 2, 2011, pp. 25-34.
[15] M. Endsley, “Toward a Theory of Situation Awareness in
Dynamic Syst ems,H uman Factors, Vol. 37, No. 1, 1995,
pp. 32-36.
[16] C. Gutwin and S. Greenberg, “The Importance of Aware-
ness for Team Cognition in Distribut ed Collaboration,In:
E. Salas and S. Fiore, E ds., Team Cognition: Understand-
ing the Factors That Drive Process and Performance,
American Psychological Association, Washington DC,
2004, pp. 177-201.
[17] A. Pokahr and L. Braubach, “The Active Components
Approach for Distributed Systems Development,” IJ-
PEDS, Vol. 28, No. 4, 2013, pp. 321-369.
[18] Distributed Management Task Force, Inc. (DMTF),
Common Information Model (CIM) Specification, Ver-
sion 2.2,” 1999.
[19] J. H. Zhang, Z. Z. Shi, J. P. Yue, et al., “Joint Intention
Based Collaboration,” IJCAI-13 Workshop WIS2013,