Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Resource Coordination in Single Agent and Multiagent Systems

Gifty Edwin and Michael T. Cox


Department of Computer Science and Engineering
College of Engineering & CS
Wright State University
Dayton, OH 45435-0001
{egifty;mcox}@cs.wright.edu

Abstract information, handling it becomes difficult. The third reason is


that due to inherent parallelism in MAS, the time required for
An intelligent multiagent system includes a number of a multiagent system to come up with a solution maybe less
software agents interacting to solve a problem. Various than the time required by a single planner with the same goal.
agents are responsible for planning different tasks. There is considerable delay in order to achieve efficient
These tasks must be coordinated to solve complex coordination. Scalability, robustness and nature of the domains
problems. A number of reasons exist for which are some other reasons for using MAS.
coordination among the agents is necessary, and
numerous issues have to be tackled to achieve efficient Our research concentrates on two types of coordination,
coordination. The focus of this paper is resource namely resource coordination and goal interaction
coordination. An agent requests a resource from another coordination [4]. Resource coordination is concerned with the
agent when it does not have enough resources to intelligent allocation of resources among the different agents
construct a complete plan. We present a cost-benefit and goals. An agent that is planning may find that it cannot
analysis that can be performed to determine if the achieve one of its goals because of a lack of a resource. The
requesting agent receives the requested resource. We problem faced here is how to determine if an agent needs a
present empirical results that demonstrate the relative resource. If so, where does it get that resource? This agent can
goal satisfaction achieved in the PRODIGY planner with request the resource from another agent or reassign a resource
and without resource coordination. We have from one of its other goals, if there are any. An agent can also
implemented a preliminary multiagent system called get resources from the environment if new resources become
COMAS (CoOrdination in MultiAgent Systems) that available. Another problem is, what happens if the resource
implements resource coordination. request is not granted? We propose that if a resource request is
not granted, the goal is minimally altered. For example, if the
city cannot be captured, then capture the local airbase.
Intelligent resource allocation among the agents in multiagent
1: Introduction systems results in better and efficient overall plans for the
individual agents. This is mainly because of higher goal
In multiagent systems (MAS) several autonomous agents
satisfaction that is achieved with resource coordination.
interact to solve complex tasks. Each MAS agent has its own
Resource coordination is required not only in multiagent
goal. These goals must be coordinated to achieve better and
systems, single agent systems with multiple goals need
more efficient overall plans. Another important characteristic
resource coordination. In this case, a resource is reassigned
of MAS is that the environment is no longer static. As
from one of its other goals to the goal that needs a resource.
opposed to single agent systems, the dynamics of the
environment are affected by the actions of other agents in that
Section 2 describes the resource coordination. It describes
environment. An example domain where MAS could be used
single agent resource contention and multiagent resource
is a system planning in a military scenario. Different agents
contention. It also describes a cost-benefit analysis that can be
plan for different goals such as deploying units, destroying
done to determine if a resource request is to be satisfied.
enemy units and securing rivers. A number of reasons exist for
Section 3 describes the implementation of a preliminary
using MAS. The first and main reason for using MAS is that it
system called COMAS [4]. Section 4 describes the empirical
helps us understand teamwork and social behavior in a human
results of COMAS planner with and without resource
organization. Secondly, a planner with multiple goals needs
coordination. Finally, section 5 is the conclusion.
the information of the resources available and the states
associated with each of its goals. Since there is a deluge of
2: Resource Coordination “river-impassable” to the goal “restricts-movement-across-
river”.
While planning, an agent may find that it cannot achieve a
goal because of a lack of resources. To handle such a situation, 2.2: Resource Contention Problem
an agent has at least two competing options over plan failure.
First, this agent can request a resource from another agent or Consider the following problem (Bridges Problem) from the
transfer a resource in service of another goal if any exist. It air campaign-planning domain taken from [3]. The mission is
could also get a resource from the environment, if new to make two rivers, R1 and R2, impassable by removing the
resources become available. If the agent gets the resource, it crossings over them. The difficulty is that R1 has 2 bridges
can achieve the originally intended goal. Second, the agent (railroad and highway) but only 1 F-15 and R2 has 3 bridges
may change its goal to a similar one that is achievable with the (railroad, highway and local road) and 2 F-15s. Figure 1
resources it already has. The main reason for resource shows the situation. Thus, 3 F-15s constitute insufficient
coordination, as mentioned earlier, is to achieve better and resources to achieve the overall goals. This is insufficient due
more efficient overall plans. Resource coordination is required to the following assumptions. First, an F-15 fighter can
not only in multiagent systems. Single agent systems with destroy only 1 bridge and damage any number of additional
multiple goals also need resource coordination (i. e. bridges. Second, a destroyed bridge has 100% reduction in
reallocation). transportation over it, whereas, a damaged bridge has φ%
reduction in transportation over it.
This section is divided into 4 subdivisions. In the first
subdivision, the concept of a goal transformation and the In this example, we consider the following question. If an
necessity for it is discussed. The second subdivision details the extra F-15 becomes available, where should it be assigned?
cost-benefit analysis that is used to determine the resource
Assuming φ is 50% and goal satisfaction is defined as the total
allocation. The third subdivision gives an example of a
reduction in transportation capacity over the whole river. If the
resource contention problem and describes how efficient
F-15 is assigned to R1, we get 100% (both the bridges across
resource allotment can result in better plans. The fourth
the R1 are destroyed) goal satisfaction for R1 and 83% (2
subdivision discusses resource coordination as achieved in a
bridges are destroyed and 1 bridge is damaged, 250/300 *100)
single agent system. The final subdivision deals with resource
goal satisfaction for R2. If it is assigned to R2, we get 75%
allocation in a multiagent system.
goal satisfaction for R1 and 100% goal satisfaction for R2.
The assignment of the extra F-15 to R1 is therefore preferable
2.1: Goal Transformations because the overall goal satisfaction is optimized.
In a dynamic environment, the world changes during planning
either by the actions of agents or by exogenous events.
Therefore, goals may become obsolete. For example, a goal of
securing city1 may become obsolete, if all enemy units move
to city2. The planner, being aware of the world changes,
minimally alters its goal to adapt to the change. That is the
goal is changed to securing city2.

A goal transformation [3] is defined as a movement of a goal


in a goal space and implements the concept of goal change.
The two reasons for using goal transformations are as follows.
First, the planner may find that it does not have enough
resources to achieve its goal. This use of a goal transformation
is relevant to resource coordination. Second, the planner may
sense changes in the environment that dictate an adjustment in
the planning process. The relevance of this second type
becomes clear in the next section.

The advantage of using goal transformations in planning is


obvious when no plan is possible for the original goal. In this
case, a minimal goal shift is clearly warranted. An example of
a goal transformation is an operationalization. Replace a vague
goal of making a river impassable with concrete goals to
destroy each bridge across it. An example of an erosion Figure 1: Resource contention problem.
transformation would be to change the goal predicate from
2.3: Cost – Benefit Analysis Given:
• nfi = number of fully achieved goals of the requested
The resource contention problem examines the general agents
problem of where to assign such extra resources, if they • npi = number of partially achieved goals of the requested
become available. In this subsection we consider an agent that agents.
requests a resource from another agent rather than the case • 0 < i < number of agents, i <> A.
where the resource becomes available exogenously. We • Pi = priority assigned to ith requested agent.
present a simple cost-benefit analysis that can be performed to
determine whether the resource transfer can be made, or the
requesting agent changes its goal, because it does not have = min ∪i = 1..n, i <> A Pi * {(100 nfi + φnpi)
enough resources. There are ‘N’ number of agents, each agent – [100(nfi -1)+ φ(npi + 1)]} * 100
making a request for a resource from N-1 other agents. The
Cost-Benefit analysis from the point of view of requesting 100nfi + φnpi
agent ‘A’ is stated as
• The requesting agent’s Goal Benefit is the percentage Assuming all the priorities are equal and by using simple
ratio of improvement in performance to performance with mathematics this equation can be reduced to
resource. The improvement in performance is the difference
between performance with resource and without resource. 100 - φ
• Requesting agent’s Goal Cost is value of the loss of the
requested Agent (or the loss incurred with respect to another
goal when reassigning its own resources). 100nfA + φnpA
• If ((Goal Benefit minus Goal Cost) is greater than 0), then
grant resource request (or make the reassignment), if not, then which is the minimum difference between the goal satisfaction
perform transformation (e.g., goal erosion) on the requesting if the resource request is not granted and the goal satisfaction
goal. possible if the resource request was granted.
We call the benefit minus cost value Utility. The utility value
is calculated using a simple formula. In the Bridges Problem, assume river R1 has 1 resource and
R2 has 3 resources. If R1 makes a request for a resource, the
Given: cost-benefit analysis is done as follows, assuming the
• φ = transformation factor (a measure of goal satisfaction priorities are 1.
in a partially achieved goal). It is assigned a value between 1 nfA = 1, npA = 1, nfi = 3, npA = 0,  φ = 50, PA = 1, P1 = 1.
and 1001. Therefore,
• nfA = number of fully achieved goals of the requesting Goal Benefit = [100 * (1 + 1) + (50 * 0)] – [100 * 1 + 50 * 1]
agent / [100 * (1 + 1) + (50 * 0)] * 100
• npA = number of partially achieved goals of the = (200 –150) / 200 * 100
requesting agent. = 25
• PA = priority assigned to the requesting agent’s goals.
Goal Cost = [(100 * 3) + (50 * 0)] – [(100 * 2) + (50 * 1)]
/ [(100 * 3) + (50 * 0)] *100
Goal Benefit = PA *{[(100(nfA + 1) + φ (npA– 1)] = (300 – 250) /300 * 100
– (100nfA +φ npA)}*100 = 16.66

Utility = 25 – 16.66
100(nfA + 1) +φ(npA – 1) = 8.33

The previous section was about multiagent choices. The next


Assuming all the priorities are equal and by using simple
section examines the choices in a single agent system.
mathematics this equation can be reduced to
100 - φ 2.4: Single Agent Resource Contention

100nfA + φnpA + 100 - φ We have solved the Bridges problem using a single copy of
the PRODIGY planning and learning architecture [1][6] agent
1
50 in the Bridges Problem because of the assumption that
damaged bridges are reduced in transportation capacity by 50
%.
whose goals are to make R1 and R2 impassable. PRODIGY
was designed and implemented at Carnegie Mellon
University. It employs a nonlinear state-space planner and
follows a means-ends analysis backward-chaining search
procedure that reasons about both multiple goals and multiple
alternative operators from its domain theory appropriate for
achieving such goals. A domain theory is composed of a
hierarchy of object classes and a suite of operators and
inference rules that change the state of the objects. A planning
problem is represented by an initial state (objects and
propositions about the objects) and a set of goal expressions to
achieve. During search for a solution the plans considered are
called incomplete plans. In PRODIGY, an incomplete plan
consists of two parts, the head-plan and the tail-plan. The
head-plan is achieved by applying a sequence of operators to
the initial state. The tail-plan is built by a partial-order
backward-chaining algorithm, which starts with a goal
statement and adds operators, one by one to achieve
preconditions of other operators, that are not true in the current
state. an example of backward chaining is that to deploy an air
unit to a forward base, base must be secured. There is a gap Figure 2: The decision point, to grant a resource
between the head and the tail. The purpose of planning is to request or perform goal transformation.
bridge the gap. Planning decisions consist of choosing a goal
from a set of pending goals, choosing an operator (or inference 2.5: Multiagent Resource Contention
rule) to achieve a particular goal, choosing a variable binding
for a given operator, and deciding whether to commit to a Planning for the Bridges Problem using a single agent system
possible plan ordering and to get a new planning state or to has all the disadvantages of using a single agent system where
continue subgoaling for unachieved goals. Different choices multiagent systems can be used amicably. To solve the
give rise to alternative ways of exploring the search space. The Bridges Problem using multiagent systems, we assign the goal
search is guided by what are called the control rules. of making R1 impassable to one agent (A1) and the goal to
make R2 impassable to another (A2). To understand the
The problem that was given to PRODIGY was to make rivers resource contention in multiagent systems, consider the same
R1 and R2 impassable by destroying the bridges across them. problem solved in the case of single agent systems, that is R1
River R1 has two bridges and one F-15 fighter assigned to it. has 2 bridges and 1 F-15 and R2 has 3 bridges and 3 F-15s.
River R2 has three bridges and three F-15 fighters. This
problem is different from the resource contention example in During planning A1 finds that it needs an F-15 to achieve its
which, R1 had 1 F-15 and R2 had 2 F-15s assigned goal. It makes a request for an F-15 from A2. As mentioned
respectively. Moreover, previously the resource becomes earlier, there is a cost associated with transferring a resource,
available during the course of planning. In this case, during which is the servicing agent’s (A2’s) loss (Value of achieving
planning, PRODIGY finds that it does not have a resource to goal less value of achieving the transformed goal) this cost is
achieve one of its goals, in this case R1, then, it deliberates as sent to A1, which in turn does the Cost-Benefit analysis to see
to changing its goals or reassigning resources. During if the utility value is greater than zero. Because the utility is
planning, we find a point where we know that to achieve the greater than zero, the agents agree for a resource transfer.
pursuing goal a resource is required. We can find this point After the transfer, if A2 does not have enough resources it
using control rules that are part of PRODIGY. Now the Cost- may perform a goal transformation or may request a resource
Benefit analysis is performed to determine if the requesting from another agent in the environment. If the utility lesser than
goal gets the resource. As mentioned, the resource could be zero, A2 does not let go of its resource, A1 may either lower
from the environment. The cost-benefit calculation helps goal expectations or may request for the resource from another
determine if the extra resource is assigned to this goal or to the agent if there is one. In this case, the utility value is greater
other goal that needs it more. If there are no new resources and than zero; therefore, the resource is transferred from A2 to A1.
the cost-benefit analysis determines that the requesting goal 
must have the resource then, a resource transfer is made from 3: COMAS (CoOrdination in MultiAgent
one goal to the one that needs it most. Figure 2 shows the
Systems)
decision point where the Cost – Benefit analysis is done. R1
gets the resource and therefore achieves the intended goal. R2 We have implemented a preliminary Multiagent system called
had to lower goal expectations as it does not have enough COMAS (CoOrdination in MultiAgent Systems), which
resources to achieve the intended goal.
implements resource coordination. COMAS have ‘N’ number
of agents. We have used a version of PRODIGY called
Prodigy/Agent [4]1 to represent multiple planning agents.
Each agent can make a resource request to the others if it
needs a resource so as to achieve resource coordination. Each
agent passes the state change information to the other agents
after it has performed an action to achieve goal interaction
coordination.

Figure 4 shows the virtual communication for resource


requests in COMAS. Here Agent1 makes a resource request to
Agent2. The cost-benefit analysis is performed to see if the
benefit it could get if it gets the resource is greater than the
cost. Agent2 responds by sending the cost associated with that
resource. If so, Agent1 sends accept message else sends a
reject message. If Agent2 receives accept message, it gives up
its resource and performs a goal transformation, if it cannot
reach its goal. If Agent1 does not send an accept message it
does not get the resource and, in turn, does a goal Prodigy/Agent1
transformation.

Figure 4: Resource coordination as achieved in COMAS

Figure 5 shows COMAS solving the Bridges Problem. The


goal is to make river R1 and river R2 impassable. Here the R1
has 1 F-15 assigned to it and needs 1 more to achieve its goal
of making R1 impassable and R2 has 3 F-15s assigned to it.

1
The Prodigy/Agent software has taken the Prodigy4.0
planning architecture and combined it with a wrapper so that it Prodigy/Agent2
behaves as an independent agent performing a plan server
function. It uses the KQML communication language to build
communication protocols that support requests for plans (over
sockets) to achieve input goals. The software is publicly
available at www.cs.wright.edu/~mcox/ProdigyAgent. Further
details are available in Cox, Edwin, Balasubramanian, and Figure 5: An example of COMAS solving a resource
Elahi (2001). contention problem.
4: Evaluation The graph in figure 6 shows COMAS system with and without
resource coordination. First, COMAS was run on 300
planning problems without resource coordination. In this case,
We conducted an experiment to compare the relative the resources were assigned based on the first come first
performance of COMAS in terms of overall goal satisfaction served principle, which is the order of the goals in this case.
with and without resource coordination. We allowed the Second, COMAS was then run with resource coordination on
COMAS system to run first on problem files with resource the same set of planning problems. We calculated the goal
coordination. Then ran the PRODIGY system to run on satisfaction for each of the problems. The way resource
problems without resource coordination. We calculated the coordination achieved was through a Java application that
overall goal satisfaction in each case. The overall goal took the planning problem, parsed through it, and made
satisfaction is defined as the total reduction in transportation intelligent resource assignment to the goals in this case to the
capacity over the all the rivers. rivers. We then ran Prodigy/Agents on these processed
planning problems and calculated the goal satisfaction
In this experiment, the goals to be achieved are kept constant achieved.
at ten (i. e., 10 rivers to make impassable) and the number of
bridges across the rivers is randomized to be between one and The figure 6 also shows plots for worst case and best case
three. To measure the relative performance with and without scenarios. The best-case scenario occurs when there are 10
resource coordination, the number of resources available was rivers, and each river has just one bridge across it. When the
varied from 1 to a maximum of 30 and the corresponding goal number of resources is varied from 1 to 30, the first 10
satisfaction with and without resource coordination is assignments results in the best assignment and when the
calculated. Each time, the number of available resources is number of resources is 10 all the rivers are made impassable
incremented, ten different problems are generated, that is, the and therefore even if the available resources are increased the
number of bridges over the rivers is varied and the number of goal satisfaction is not increased. The worst-case scenario
resources held constant. The final goal satisfaction achieved occurs when there are 10 rivers and each river has 3 bridges
with the available resources is the average of the goal across. The total number of bridges is 30 and the number of
satisfaction from the ten different problems. The total planning resources required is 30. In this case, any combinations of
problems therefore amount to 300. resource assignment result in the same goal satisfaction. For
relative goal satisfaction achieved in PRODIGY with and
As specified in section 2, the following assumptions are made: without goal transformation see [3].
First, an F-15 can destroy one bridge and damage all the other
bridges. Second, the destroyed bridge is assigned 100%
reduction in transportation capacity, and third a damaged 5: Conclusion
bridge is assigned 50% reduction in transportation capacity.
The focus of this paper is on resource coordination in
multiagent systems. From our research, we conclude resource
Best Case coordination achieve efficient and better plans. The cost-
With Resource Coordination
Without Resource Coordination benefit analysis that we introduced is efficient, but in real
Worst Case world planning, it should include other parameters like
100 individual sub goal priorities and transportation costs in the
90 cost-benefit function. Although the cost-benefit function is
80
simple, it demonstrates that an intelligent choice of whether to
get a resource from another agent or to change the goal of an
70
agent will lead to higher quality plans that have a higher
60
overall goal satisfaction. The cost-benefit analysis provides a
Utility Value

50 way to determine if a resource request can be granted.


40

30 When the requesting agent does not get the requested


20
resource, and the originally intended goal is not possible
because of that, the agent can achieve a partial goal by using
10
the concept of goal transformation. Using an efficient resource
0 allocation algorithm results in achieving a higher overall goal
11

13

15

17

19

21

23

25

27

29
1

No of Resources satisfaction on average. The experimental results in the


previous sections confirm this. An improvement in the overall
goal satisfaction is certainly warranted in real world planning.
Figure 6: Shows the relative performance with and
without resource coordination. It also is better to involve the human while making these
decisions in the loop. A mixed-initiative approach, where the
decision making capabilities of the human agent is taken into
account, is thus a better way for making these critical
decisions. By using the cost-benefit analysis, the user can
make better decisions than making decisions without the cost-
benefit analysis.

6: Acknowledgements
This research is funded by Dayton Area Graduate Studies
Institute (DAGSI) under grant #HE-WSU-99-09, and the Air
Force Research Laboratory (AFRL) under grant number
F49620-99-1-0244 and by the Information Technology
Research Institute (ITRI). We thank Boris Kerkez, Greg
Kramer and Mamun Elahi for comments on a draft of this
paper.

7: References
[1] Carbonell, J. G., Blythe, J., Etzioni, O., Gil, Y., Joseph, R., Kahn,
D., Knoblock, C., Minton, S., Pérez, A., Reilly, S., Veloso, M. M.,
and Wang, X. 1992. PRODIGY4.0: The Manual and Tutorial,
Technical Report, CMU-CS-92-150, Computer Science Dept.,
Carnegie Mellon University.

[2] Cox, M. T., Edwin, G., Balasubramanian, K., & Elahi, M. 2001.
MultiAgent Goal Transformation and Mixed-Initiative Planning
using Prodigy/Agent. Proceedings of the 5th World Multiconference
on Systemics, Cybernetics and Informatics (SCI 2001).

[3] Cox, M. T., & Veloso, M. M. 1998. Goal Transformations in


Continual Planning. In M. desJardins (Ed.), Proceedings of the 1998
AAAI Fall Symposium on Distributed Continual Planning, 23-30.
Menlo Park, CA: AAAI Press / The MIT Press.

[4] Edwin, G. & Cox, M. T. 2001. COMAS: CoOrdination in


MultiAgent Systems. Proceedings of the 5th World Multiconference
on Systemics, Cybernetics and Informatics (SCI 2001).

[5] Veloso, M. M., Pollack, M. E., and Cox, M. T. 1998.


Rationale-Based Monitoring For Continuous Planning in Dynamic
Environments. In R. Simmons, M. Veloso, and S. Smith, eds.,
Proceedings of the Fourth International Conference on Artificial
Intelligence Planning Systems, 171-179. Menlo Park, AAAI Press.

[6] Veloso, M. M., Carbonell, J., Perez, A., Borrajo, D., Fink, E., &
Blythe, J. 1995. Integrating planning and learning The PRODIGY
architecture. Journal of Theoretical and Experimental Artificial
Intelligence. 7(1): 81-120.

You might also like