Dynamic assignment of geospatial-temporal macro tasks to agents under human strategic decisions for centralized scheduling in multi-agent systems


Abstract—Problem: This paper addresses a centralized
scheduling problem in multi-agent systems in which the incident
commander (IC) of a disaster-response team aims to coordinate
the actions of the field units (rational agents) to minimize the
total operation time in uncertain, dynamic, and spatial
environments.
Objective: The purpose of this paper is to propose an
intelligent software system that assists the IC in the dynamic
assignment of geospatial-temporal macro tasks to agents under
human strategic decisions. This system autonomously executes a
heuristic algorithm to minimize the maximum total dependent
duration according to human high-level strategies.
Results: The result is a schedule composed of macro decisions,
each comprised of seven types of information: 1) what task type
is going to be accomplished, 2) who (a subset of agents) are
assigned to this assignment, 3) where this task is to be performed
(a road segment or zone as a macro geospatial object) containing
a subset of tasks, 4) when operations start, 5) when operations
finish, 6) how many tasks are estimated to be completed, 7) what
task types and the estimated number to be revealed (identified”
or “enabled) in this location to complete this job.
Conclusion: This result, which is a feasible solution for the
addressed problem, permits the IC to coordinate agents,
partially specify the activities of the agents in time and space,
minimize the overall execution time for all the tasks, calculate
the correct time to revise the strategic decisions, evaluate the
efficiency of the high-level strategy, and estimate the makespan.
Index Terms—Task assignment, centralized scheduling,
multi-agent systems, macro tasks, disaster emergency response,
heuristic algorithm, coordination, incident commander
I. INTRODUCTION
The objective of scheduling is to minimize the overall
execution time for all tasks by properly allocating them to
processors/machines without violating the precedence
constraints among the tasks [1]. The input to these problems is
usually a directed acyclic graph (DAG) that provides
precedence, dependency, priority among tasks, cost, and other
information for the tasks [2]. The total run-time is called the
makespan.
Manuscript received August 13, 2013; revised November 20, 2013.
Reza Nourjou is with the Informatics Graduate School and DPRI, Kyoto
University, Japan (e-mail: nourjour@imdr.dpri.kyoto-u.ac.jp).
Stephen F. Smith is with the The Robotics Institute, Carnegie Mellon
University, USA (e-mail: sfs@cs.cmu.edu).
Michinori Hatayama is with the Disaster Prevention Research Institute,
Kyoto University, Japan (e-mail: hatayama@imdr.dpri.kyoto-u.ac.jp).
Norio Okada is with the School of Policy Studies, Kwansei Gakuin
University, Japan (e-mail: okada-n@kwansei.ac.jp).
Pedro Szekely is with the Information Sciences Institute, University of
Southern California, USA (e-mail: pszekely@isi.edu).
Urban search and rescue (USAR) is considered the major
function of disaster emergency response operations. The
objective of USAR is to reduce the number of fatalities in the
first days after a disaster [3]. Scheduling is a crucial issue for
incident commanders (ICs) who must manage and coordinate
the activities of the field units (rational agents) during the
disaster emergency response. It aims to identify and schedule
the actions of the agents by appropriately assigning and
allocating their tasks. Although scheduling enables the agents
to accomplish tasks in the minimum overall time, it is difficult
to solve this problem under uncertain, spatial, and dynamic
situations.
A responder team such as the INSARAG (International
Search and Rescue Advisory Group) has a hierarchical
structure consisting of two levels. The lower level in the
hierarchy includes the rational agents. They are spatially
distributed in a geographic environment. They perceive their
local environment, execute strategic decisions made by the
top node, rationally make decisions regarding their own
actions, coordinate their actions with each other in a
distributed approach, execute tasks, and report to the
operation center [4]. An IC, located in the top node, has a
global and timely view of the entire disaster situation. His role
is to coordinate and command the agents using strategic
planning and scheduling in a centralized approach. As a result,
a team is a society of cooperative agents that attempt to
maximize the global utility of the team.
Coordination is the act of managing interdependencies
among activities [5]. Coordination in disaster emergency
response includes the management of task flow, recourse,
information, decisions, and responders [6]. This paper
recognizes the importance of five main types of coordination
in a disaster-response team:
1) Task Dependencies: The Enabling dependency between
tasks specifies that when a task is completed, it triggers
the initiation of other, dependent tasks. In other words,
the time that a disenabled task becomes enabled depends
on the finishing time of the task to which it is dependent.
Agents can carry out only tasks whose state is enabled.
2) Action Dependencies: Heterogeneous agents can
perform tasks that have been previously revealed by the
actions of other agents. Some agents may wait until their
tasks are released by others.
3) Redundant Actions: Multiple agents may possess
overlapping capabilities that allow them to be involved in
the same tasks. This can cause conflict between actions or
redundant actions.
4) Information Sharing: Information sharing allows
Dynamic Assignment of Geospatial-Temporal Macro
Tasks to Agents under Human Strategic Decisions for
Centralized Scheduling in Multi-Agent Systems
Reza Nourjou, Stephen F. Smith, Michinori Hatayama, Norio Okada, and Pedro Szekely
International Journal of Machine Learning and Computing, Vol. 4, No. 1, February 2014
39DOI: 10.7763/IJMLC.2014.V4.383

agents to know 1) the current state of the tasks, 2) what
tasks are going to be performed, by whom and when, and
3) what tasks are estimated to be revealed when, and
where. These data enable an IC to coordinate the agents
and facilitate multi-agent scheduling.
5) Agent Allocation: Agents are considered as limited
resources or machines that must be optimally scheduled
and allocated to tasks.
Many good approaches have been developed to address the
scheduling problem. However, there remain challenges to the
problem that are addressed by this paper. We categorize these
limitations and deficiencies as follows:
1) These approaches explicitly and fully schedule and
specify actions for the rational and autonomous agents
who are located in uncertain and dynamic environments.
They attempt to associate a specific task with a specific
agent. They do not provide an efficient approach for an
IC because situational awareness that an IC has from the
global picture is different from the perception that the
agents have from their local environment.
2) They are automated systems that do not involve a human
in the loop. The primary obstacle is the scale of the
operation. It is currently unrealistic for a fully automated
system to effectively determine all of the possibilities that
may arise during the execution of the tasks in a complex
environment.
3) They do not address the scheduling problems in which
macro tasks are spatially distributed in a complex
environment. Moreover, geospatial reasoning and
geo-information analyses play a significant role in
problem solving. Spatiotemporal macro tasks and
characteristics of this environment make the assignment
problem more difficult.
This paper addresses the capability of GICoordinator in
macro action scheduling. The IC is equipped with a computer
that runs an instance of GICoordinator. GICoordinator is a
GIS (Geographic Information Systems)-based intelligent
software system that collaborates with an IC (as a human) and
assists the IC in the action coordination of agents in complex
situations such as disaster emergency response [7]. This
system autonomously executes an algorithm for the
assignment of spatiotemporal macro tasks to rational agents
under human strategic decisions in a spatial and dynamic
environment. The goal of this algorithm is to minimize the
maximum total dependent duration by the dynamic
assignment of idle agents to these tasks.
The contribution of this paper is a heuristic algorithm that
calculates a feasible solution for the addressed problem. This
algorithm enables us to develop a software agent that can
collaborate and cooperate with an IC on scheduling in
multi-agent systems for agent coordination. It calculates a set
of macro (semi-strategic) decisions that temporally constrain
the actions of the agents according to the human strategy.
II. LITERATURE REVIEW
The scheduling problem for multiprocessor systems is
stated as "How can we execute a set of tasks T on a set of
processors P subject to some set of optimizing criteria C? " A
taxonomy of scheduling algorithms in parallel and distributed
systems was presented by [8]. Algorithms have been
developed to solve this problem [1, 9].
The algorithm presented by [10] aims to minimize the
maximum lateness and total delay for multi-machine
problems concerned with the scheduling of single-operation
jobs. Several heuristic algorithms were proposed by [11, 12]
for the assignment of n independent tasks to m unrelated
parallel machines to minimize the maximum completion time.
This approach was extended by this paper to propose a new
algorithm for the addressed problem.
To assign spatially distributed tasks to field units (agents)
in USAR management, algorithms were suggested by [4, 13].
These algorithms involved geographic information and
geospatial analyses in the effective assignment of tasks to
agents.
There are several mechanisms, techniques, and algorithms
for addressing task assignment to agents (humans, robots, or
intelligent systems). They assist agents to achieve their
objectives and to maximize the benefits of the system. There
are works that address the task-scheduling problem in
multi-agent systems [14, 15], multi-robot systems [16],
disaster-emergency teams [17], robocup rescue simulations
[18], and strategic decision making for coordinating actions
of a USAR team[19].
III. PROBLEM STATEMENT
The problem addressed by this paper has characteristics
described in the following subsections.
A. Problem Domain
USAR is a significant participant in earthquake
disaster-response situations. The overriding goal is to rescue
the greatest number of people, trapped under the debris of
damaged buildings, in the shortest amount of time. USAR
tasks involve a sequence of dependent tasks: (1)
reconnaissance and assessment by collecting information on
the extent of the earthquake damage, (2) search and locate
victims trapped in collapsed structures, (3) extract and rescue
trapped victims, and (4) transport and dispatch injured
survivors to hospitals or shelters. Rescue tasks are classified
into light, medium, and heavy rescue.
USAR tasks are location-based entities that are distributed
in a geographical area. A specific task needs a specific
capability or several synchronous capabilities to complete in
an acceptable duration. Capability requirements determine
what agents are allowed to do what tasks. There is an
uncertainty in task duration, distribution of tasks, and
outcome of tasks.
B. Agents
Field units are considered mobile, spatial, rational,
heterogeneous, and semi-autonomous agents. Their principal
roles are to perform tasks suited to their capabilities. They
possess different capabilities that allow them to engage in
executing tasks for which they can provide the required
capability. Because agents perform their tasks with different
speeds, there is an uncertainty in action duration, speed, and
outcome.
40

The IC, as the planning human, attempts to coordinate the
disaster-response management. S/he has inaccessible, global,
and uncertain information regarding the state of the
environment. His observation and perception of a disaster
situation is different from that of the agents in their local
environment.
C. Macro Tasks
Macro task information forms a global picture of the task
environment for ICs. A macro task is the aggregation of all the
tasks that 1) are from a same task type and 2) are spatially
contained within a specific geographic object or adjacent to a
road segment in a definite time. Macro tasks are encoded by
road segment in this paper. Road segments are surrounded by
buildings that contain USAR tasks. Therefore, a segment
contains a number of macro tasks that are located near this
geographic object.
A macro task contains two variables: 1) an enabled number
of tasks, and 2) a disenabled number of tasks. The enabled
variable indicates how many homogeneous tasks are observed,
discovered, or revealed within an associated geo-object.
Agents can only undertake tasks that are enabled. The
disenabled variable indicates how many homogeneous tasks
are estimated to be revealed or discovered in the future. They
have dynamic and uncertain quantities that may vary over
time because agents can complete an enabled task while other
agents may reveal (enable)new ones. These numbers provide
an estimation of the total duration and total capabilities that
are required to complete this macro task.
There is a complex, interdependent relationship among
variables of macro tasks that are associated with a geo-object.
For example, a whole reconnaissance macro task (enabled
number plus disenabled number) can reveal only the
disenabled part of a search macro task.
D. Human Strategic Decisions
A high-level strategy specification enables an IC to express
his intuition and initiative for solving multi-agent planning
problems. A strategy decomposes a complex problem into a
set of smaller problems under human supervision. A strategy
contains a set of parallel, interdependent, and prioritized
threads. A thread is a sub-problem that is composed of a
unique ranking, a subset of agents, a subset of task types, and
a subset of a geographic area [19].
Strategic decision-making is the problem of the distribution
of agents among the threads in an explicit time. Because
agents may be shared among threads, an agent should be
assigned to only one thread at a time. This agent will be
available for the next thread whenever s/he is no longer
required and is released by the thread. A human strategic
decision defines what agents are assigned to what threads in
an explicit time. A thread constrains the actions of the
assigned agents according to its specification. A strategic
decision should be refined and adapted to new states.
E. Scheduling in Multi-Agent Systems
Scheduling is the problem of the assignment of macro tasks
to rational agents under human strategic decisions in
multi-agent systems. It does not explicitly specify actions for
the agents, however, it does partially identify actions and
dynamically constrains agent activities and behaviors in time
and space. It allows rational agents to make and adapt their
own decisions according to calculated scheduling. Scheduling
includes the characteristics as follows:
Scheduling has a macro characteristic. More than one agent
can be assigned to a macro task simultaneously. Assigned
agents form a coalition that will execute this assignment
cooperatively.
Assignments are dynamic. Over time, new agents can be
assigned to a macro task to which a set of agents have been
previously allocated.
Macro tasks retain assigned agents until the enabled tasks
are completed. Released agents can be reassigned to other
macro tasks.
The objective is to minimize the overall time for the USAR
task execution.
Because of the spatial distribution of tasks, agents need
travel time to reach macro tasks moving from one location to
another using the road network. In addition, streets have
varying states such as blocked or cleaned that affect the
shortest paths.
A coalition formed by several professional agents can do a
macro task faster than another coalition. An efficient coalition
has many capabilities and can finish a macro task faster than
another coalition.
Action scheduling is calculated with regard to a human
strategic decision as long as this strategic decision is valid. It
consists of a sequence of temporal scheduling.
Scheduling changes the state of the macro tasks. An
assignment can define how much of a macro task is scheduled
to be addressed, or it can define how much of another macro
task will be revealed after completion.
IV. APPROACH
The proposed approach shown in Fig. 1 is used by ICs for
multi-agent coordination. The focus of this paper is on the
fourth component.
Fig. 1. Role of the central scheduling problem (the 4th
component) in the
proposed approach for strategic planning and scheduling in disaster
response team [20].
41

Whenever a human strategic decision is made for solving a
problem, this system autonomously executes a heuristic
scheduling algorithm for the dynamic assignment of
spatiotemporal macro tasks to rational agents under human
strategic decisions. The algorithm, which is shown in
Algorithm 1, is composed of 1) select efficient agents, 2)
select active macro-tasks, 3) identify the release time, 4)
select idle agents, 5) nominate macro task, 6), calculate
utilities, 7) find the highest utilities, 8) assign agents to the
macro tasks, and 9) calculate the earliest finish time.
This paper applies the SAP data model [20] for the
presentation and formulation of the problem. It enables us to
design, implement, and develop the proposed algorithm. Fig.
2 shows a part of this data model that presents the important
information classes.
A. Select Efficient Agents
This algorithm aims to select efficient agents for the macro
tasks. It has two purposes. The first is to identify what agents
can be assigned to what macro tasks. The second is to
calculate the maximum ability that an assigned agent can
provide to a macro task. The selection of the agent for the
macro task is important because selecting an agent states that
this agent is a useful candidate who can be legally assigned to
and perform this macro task. This algorithm considers four
key criteria to efficiently select agents for a definite macro
task:
1) Does this macro task contain any task?
2) Does the human strategic decision allow this agent to be
allocated to the location of this macro task?
3) Does the human strategic decision allow this agent to be
assigned to the task type of this macro task?
4) Does this agent have any useful capability that is required
by this macro task and what maximum capability is
provide by this agent?
An agent can be selected for more than one macro task and
more than one agent can be selected for a macro task. After
agent selection for a definite macro task, this algorithm
creates a set of legalAssignment elements for this macro task.
A macro task may contain a number of legalAssignment
elements that define what agents are allowed to be assigned to
this macro task. A legal assignment is composed of three
kinds of information: 1) the Id of the agent, 2) a maximum
capability that the associated agent provides with regard to the
capability requirements of this macro task, and 3) an
estimated time when this agent will arrive in the location of
this macro task and will be available to begin.
An agent is efficient for a task if s/he provides any
capability required by the task under the human strategic
decisions. To simplify this algorithm, we assumed that each
task type requires a specific capability and agents provide
Fig. 2. Part of the SAP data model that includes important information
classes for this paper [20]Algorithm 1. Heuristic algorithm for dynamic assignment of
spatiotemporal macro tasks to rational agents under human strategic
decisions for centralized scheduling in multi-agent systems
42

different capabilities. This assumption ignores tasks that may
require simultaneous capabilities. This algorithm must also
consider not only the enabled tasks but also the disenabled
tasks. These tasks are estimated and expected to be revealed,
enabled, or discovered by other tasks in the future. Because of
the dynamic environment, this algorithm enables the system
to automatically and timely recognize agents who are no
longer efficient by macro tasks e.g. there is no task for these
agents to carry out for this thread.
B. Select Active Macro Tasks
The aim of this algorithm is to select active macro tasks
from the task environment. It is important to identify and
recognize only a subset of the macro tasks to which agents can
be potentially allocated. Agents can only work on enabled
tasks. If agents are assigned to disenabled tasks, they become
inactive and must wait until tasks are discovered, revealed, or
enabled by completing other tasks.
This algorithm applies two criteria to identify a macro task
as active:
1) Does this macro task contain any enabled tasks?
2) Does this macro task contain any legalAssignment?
C. Select Idle Agents
This algorithm aims to select idle agents from the team at
the current time, g2. Selected agents are candidates who will
be assigned accordingly to other active macro tasks.
There are two sources that dynamically and temporally
provide idle agents for this algorithm:
1) IC who defines and enters new agents to the team.
2) Macro tasks that release assigned agents.
A macro task releases its agents assigned to enabled tasks
by the time g2. This algorithm accesses the information from
the macro tasks and agent allocation and identities all idle
agents who have been released by a macro task at this time. A
legalAssignment element is used to encode a released agent
and present a last location (Id of a road segment). Finally, the
algorithm adds the selected agents to the current collection, L,
in the algorithm.
D. Identify the Release Time
This algorithm determines whether it is a correct decision
to release a subset of idle agents at time g2. It is important, in
a timely manner, to adapt a strategic decision to a new
situation by releasing unneeded agents from threads and
sending (entering) them to other threads that may require
them. This algorithm helps the system to select a choice
between: 1) continue scheduling with all the idle agents under
the same strategic decision, or 2) stop scheduling and let the
system revise the current strategic decision.
As Algorithm 2 shows, this algorithm considers two
criteria to keep an idle agent for task assignment:
1) Is there any macro task for which this agent has been
categorized as an efficient agent previously?
2) Is it estimated that an identified macro task release any
new task (enabled tasks or disenabled tasks) at the current
time?
E. Nominate Macro Tasks
This heuristic algorithm aims to select a subset of the active
macro tasks. Idle agents are assigned to the nominated macro
tasks. Unlike a search algorithm, this algorithm selects the
best node and extends it to reach the goal node at which all the
tasks are completed.
The logic of this algorithm is to select macro tasks that have
a strong effect on minimizing the makespan. This algorithm
nominates macro tasks that have a “longest total dependent
duration”. Two sources of information are important for the
calculation of this dynamic variable for each macro task for
the current time:
1) How many enabled tasks are contained in this macro task?
Alternatively, when does the enabled portion of this
macro task complete according to the previous schedule?
2) What is the estimate of how many dependent tasks this
macro task can reveal?
F. Calculate Utilities
This algorithm aims to calculate a utility matrix of idle
agents for the nominated macro tasks. It is important to
answer the question “How much can a macro task benefit
from allocation of a specific agent?” This matrix assists in
selecting and allocating the best agents for the macro tasks.
The benefit statement asks “How much will an agent reduce
the finish time of a macro task, if this agent is assigned to the
task?” This algorithm takes into account several factors in the
calculation of the benefit of a specific agent for a specific
macro task:
1) Is it legal for this idle agent to be assigned to this
nominated macro task?
2) When will this agent start doing tasks? The dynamic
amount of this time is dependent on three factors: 1)
moving speed of this agent, 2) release time of this agent,
and 3) travel time and distance between two locations in
the road network.
3) What capability does this agent provide for this macro
task? With what speed can this agent carry out tasks?
4) How much can this agent improve performance of a
coalition that has previously been assigned to this macro
task?
Algorithm 2. “Identify_theRelease_Time()” algorithm that determines
whether it is a correct decision to release a subset of idle agents from a
thread now
43

This utility matrix is used to present a matrix of utility
elements that idle agents provide for nominated macro tasks.
Rows are assigned to macro tasks and columns are assigned to
agents. A set of numbers in a cell shows the utility information
that the relevant agent can provide for the relevant macro task.
G. Find the Highest Utilities
The goal of this algorithm is to determine if there is a subset
of idle agents who can provide a benefit-ratio higher than 10%.
This algorithm uses the utility matrix to find these agents who
can provide acceptable benefits for the nominated macro
tasks. These agents will be assigned to the related macro tasks.
If a suitable agent is not found, all nominated macro tasks will
be removed by the algorithm.
H. Assign Agents to Macro Tasks
This heuristic algorithm is shown in Algorithm 3. It aims to
select the proper agents and assign them to the macro tasks.
This algorithm uses the utility matrix and applies a minimax
algorithm to select the best agents for the best macro tasks.
This algorithm selects all agents that decrease the finish time
of the macro tasks by more than 10%. If an agent is allocated
to a macro task, the associated row and column of the matrix
is removed and the assigned agent is removed from the list of
idle agents.
I. Calculate the Earliest Finish Time
This algorithm aims to calculate the earliest time at which
agents will be released by a macro task and become idle. It
updates the problem state that includes agents, tasks,
scheduling, and the finish time of each macro task. This
algorithm enables us to simulate the problem and estimate
future states. A new subset of macro tasks is identified. The
results allow other algorithms to continue scheduling.
V. EVALUATION AND RESULTS
To evaluate the proposed approach, a USAR scenario
simulated by GIS was used. As Fig. 3 shows, a rapid response
team composed of an IC and 12 agents was assigned to the
operational area. The global objective was to accomplish,
with this team, USAR tasks in a minimum time. Table 1 shows
the strategic decision that was used for scheduling at time 0
(the initial time).
TABLE 1. HUMAN HIGH-LEVEL STRATEGIC DECISION VALID FROM TIME 0
AND USED FOR SCHEDULING
thread Id set of zones set of task types set of agents
1 zo1, zo2 T0-Reconnaissance a0, a1, a2,
a3, a4, a5
2 zo1, zo2 T1-Search a7
3 zo1, zo2 T2-SlightRescue,
T3-MediumRescue,
T4-HeavyRescue
a9, a11
4 zo3, zo4 T0-Reconnaissance,
T1-Search,
T2-SlightRescue,
T3-MediumRescue,
T4-HeavyRescue
a6, a8, a10
The two main results achieved by executing the proposed
algorithm are: 1) a feasible schedule, and 2) an
adaption/refinement time. In addition, two minor results are
achieved: 1) the makespan, and 2) the total schedule.
The schedule is composed of a number of macro decisions
that partially specify the actions of the agents. This schedule is
calculated according to the valid strategic decision shown in
Table 1. A macro decision contains seven types of
information:
1) A task type as a sub-goal that is scheduled to be
completed.
2) A set of agents who are allocated to this decision and are
responsible for performing this decision.
3) A geographic object such as a road segment that presents
the operational area of this decision.
4) A time that indicates when this decision execution starts.
Fig. 3. Two thematic maps of spatial distribution of macro tasks [20].
Algorithm 3. Agent selection for assignment of tasks to agents.
44

5) A time that indicates when this decision ends.
6) A quantity of the number of tasks that are estimated to be
completed during this decision.
7) A quantity of tasks with different types that are estimated
to be revealed at the end of this decision, at the same
location.
The adaption time is extracted from the schedule. It
estimates a correct time when the current human decision is
required to be revised by the IC. Furthermore, this time states
that this schedule ends at this time and a new schedule will be
calculated based on the strategic decision adapted.
The overall time can be calculated if the scheduling
algorithm and the planning algorithm are executed in
sequence as shown in Fig. 1. Search algorithms can apply this
method to calculate the total schedule with the estimated
makespan in order to find an optimal solution for the
scheduling problem.
Results showed that 1) the first human strategic decision
expired at time 579 (minutes), 2) the total makespan is 16859
(minuts), and 3) a sequence of twelve strategic decisions was
made. For example, Fig. 4 shows a geo-visualization of the
macro actions scheduled for an agent. An example of the task
assignment and action scheduling is presented in Fig. 5 using
another data set in order to view a better perception of the
results.
A. Discussion
This algorithm calculated a feasible, fast, and semi-optimal
solution for the addressed problem. However, the results
showed that this algorithm did not guarantee optimality of the
assignments because the task list showed that some agents
sometimes became idle. In an efficient coordination, agents
should remain active. Minimizing the makespan is dependent
on the quality of five parameters: 1) human strategy, 2)
strategic decision, 3) search algorithms, 4) scheduling
algorithms, and 5) distribution coordination among agents.
The adaption time of a strategic decision is dependent upon
the structure of a scheduling algorithm and other factors. If
several algorithms are applied to the same data, they will
reach different results. We cannot assume that the optimal
assignment minimizes the adaption time.
Tasks may include deadlines that influence maximizing the
global utility. Humans can manage this issue by specifying an
effective high-level strategy.
This paper proposed an algorithm for multi-agent
scheduling. There is a value to supporting humans by
providing a set of algorithms for the IC and allowing him to
select a suitable algorithm for each thread. The system can
apply a specific algorithm for a thread.
This paper studied a single IC. A disaster emergency
response may involve several ICs and there are
interdependencies among the team actions. It would be
valuable to study the distributed coordination among multiple
ICs and their scheduling.
This paper did not study synchronized actions, facilitating
task dependency, or resource allocations in the proposed
algorithm. These subjects could be important issues for some
domains and teams.
A comprehensive approach for multi-agent coordination
includes a loop consisting of scheduling, execution,
monitoring, adapting, and learning. Uncertain and dynamic
environments disrupt scheduling, making it necessary to
execute decisions, continuously monitor execution, and adapt
scheduling to unexpected events. Learning algorithms could
provide efficient tools to improve the efficiency and
performance of this system.
VI. CONCLUSION
This paper presented a heuristic and greedy algorithm that
appropriately assigns macro tasks to rational agents and
schedules the macro actions of agents according to human
strategic decisions. This algorithm was used to develop
GICoordinator that assists the IC in coordinating a team of
agents.
A schedule provides an IC with a feasible solution for a
coordination problem in a team. Multi-agent scheduling
partially specifies the actions of the rational agents and
constrains them temporally and dynamically. The IC
authorizes agents to autonomously make their own tactical
decisions (planning and scheduling) or adapt their activities
under these macro decisions.
The proposed system executes this algorithm while the
current strategic decision is valid. Therefore, this algorithm
calculates a time at which this strategic decision should be
adapted. The result has two advantages: 1) It enables the IC to
refine this strategic decision at the correct time, and 2) it is
used in a search algorithm that estimates a makespan,
calculates a complete solution, and finds an optimal solution
Fig. 5. Example of assignment of macro tasks to agents for another
data set
Fig. 4. 3D map that geo-visualizes macro actions scheduled for a
specific agent
45

for the multi-agent planning and scheduling.
Macro tasks enable the system to model and present the
distribution of the tasks in different geographic scales. Spatial
topology between spatial objects enables the system to extract
new views of the tasks. These tasks can be used for task
assignment. Limited only by the quality of the data, an IC can
select different geographic layers such as zones or buildings
and allocate tasks to them. The IC can specify and contain
actions in different spatial accuracies.
Information fusion algorithms extract new and useful
information from the scheduling results. They provide an
efficient situational awareness for an IC to view the state of
the global environment. They support human decisions using
tools such as geospatial reasoning, geo-visualization, and
complex queries. For example, they can answer the questions
“What activities will be completed in a specific zone during a
specific period?”, “When will a specific task be completed
and by whom?”, and “What is the task list of a specific
agent?”
Future works to improve the current contribution may
include:
1) In some cases, the tasks environment contains tasks that
may require synchronous capabilities. This means that
the actions of more than one agent must be coordinated to
provide the required capabilities for performing a
specific task simultaneously. It is important that an ideal
algorithm form a proper coalition comprised of suitable
agents and assign this group to the task.
2) Decentralized coordination of distributed schedules is a
significant issue in which multiple teams are involved in
performing the tasks. It is possible that interdependencies
between team actions exist. To maximize the joint
objective, it is necessary to apply algorithms to
coordinate the distributed schedules prepared by each IC.
ACKNOWLEDGMENT
Reza Nourjou is grateful for the financial support of
GCOE-HSE of Kyoto University. This enabled him to be a
visiting scholar at the Information Sciences Institute of the
University of Southern California and the Robotics Institute
of Carnegie Mellon University from November 2011 to
November 2012.
REFERENCES
[1] Y. K. Kwok and I. Ahmad, "On multiprocessor task scheduling using
efficient state space search approaches," Journal of Parallel and
Distributed Computing, vol. 65, no. 12, pp. 1515-1532, 2005.
[2] Y. K. Kwok and I. Ahmad, "Benchmarking and comparison of the task
graph scheduling algorithms," Journal of Parallel and Distributed
Computing, vol. 59, no. 3, pp. 381-422, 1999.
[3] F. Fiedrich, F. Gehbauer, and U. Rickers, "Optimized resource
allocation for emergency response after earthquake disasters," Safety
Science, vol. 35, no. 1, pp. 41-57, 2000.
[4] R. Nourjou, M. Hatayama, and H. Tatano, "Introduction to spatially
distributed intelligent assistant agents for coordination of human-agent
teams’ actions," in Proc. 2011 IEEE International Symposium in
Safety, Security and Rescue Robotics, 2011, pp. 251-258.
[5] W. T. Malone and K. Crowston, “The interdisciplinary study of
coordination,” ACM Computing Surveys, vol. 26, no. 1, pp. 87-119,
1994.
[6] R. Chen, R. Sharman, H. R. Rao, and S. J. Upadhyaya, "Coordination
in emergency response management," Communications of the ACM,
vol. 51, no. 5, pp. 66-73, 2008.
[7] R. Nourjou, M. Hatayama, S. F. Smith, A. Sadeghi, and P. Szekely,
“Design of a gis-based assistant software agent for the incident
commander to coordinate emergency response operations,” presented
at Workshop on Robots And Sensors Integration in Future Rescue
Information System, 2013.
[8] S. J. Chapin, "Distributed and multiprocessor scheduling," ACM
Computing Surveys, vol. 28, no. 1, pp. 233-235, 1996.
[9] Y. Wen, H. Xu, and J. D. Yang, "A heuristic-based hybrid
genetic-variable neighborhood search algorithm for task scheduling in
heterogeneous multiprocessor system," Information Sciences, vol. 181,
issue. 3, pp. 567-581, 2011.
[10] W. A. Horn, "Some simple scheduling algorithms," Naval Research
Logistics Quarterly, vol. 21, no. 1, pp. 177-185, 1974.
[11] Davis, Ernest and J. M. Jaffe, "Algorithms for scheduling tasks on
unrelated processors," Journal of the ACM (JACM) vol. 28, no. 4, pp.
721-736, 1981.
[12] C. N. Potts, "Analysis of a linear programming heuristic for scheduling
unrelated parallel machines," Discrete Applied Mathematics, vol. 10,
no. 2, pp. 155-164, 1985.
[13] A. R. Vafaeinezhad, A. A. Alesheikh, M. Hamrah, R. Nourjou and R.
Shad, "Using gis to develop an efficient spatio-temporal task allocation
algorithm to human groups in an entirely dynamic environment case
study: earthquake rescue teams," in Proc. Computational Science and
Its Applications ICCSA, Springer Berlin Heidelberg, 2009, pp. 66-78.
[14] V. Lesser, K. Decker, T. Wagner et al., "Evolution of the
GPGP/TAEMS domain-independent coordination framework,"
Autonomous Agents And Multi-Agent Systems, vol. 9, no. 1-2, pp.
87-143, 2004.
[15] E. T. O. Opiyo, E. Ayienga, K. Getao, W. O. Odongo, B. Manderick
and A. Nowe, "Game theoretic multi-agent systems scheduler for
parallel machines," International Journal of Computing and ICT
Research, vol. 1, no. 1, 2008.
[16] P. B. Gerkey and J. M. Matari, "A formal analysis and taxonomy of
task allocation in multi-robot systems," The International Journal of
Robotics Research, vol. 23, no. 9, pp. 939-954, 2004.
[17] P. Scerri, A. Farinelli, S. Okamoto, and M. Tambe, "Allocating tasks in
extreme teams," in Proc. The Fourth International Joint Conference
on Autonomous Agents And Multiagent Systems, ACM, 2005, pp.
727-734.
[18] S. D. Ramchurn, A. Farinelli, K. S. Macarthur, and N. R. Jennings,
"Decentralized coordination in robocup rescue," The Computer
Journal, vol. 53, no. 9, pp. 1447-1461, 2010.
[19] C. A. Chapman, R. A. Micillo, R. Kota, and N. R. Jennings,
"Decentralized dynamic task allocation using overlapping potential
games," The Computer Journal, vol. 53, no. 9, pp. 1462-1477, 2010.
[20] R. Nourjou, P. Szekely, M. Hatayama, M. G. Ashtiany, and S. F. Smith,
"Data model of the strategic action planning and scheduling problem in
a disaster response team," Journal of Disaster Research, 2013.
Reza Nourjou was born in Urmia, Iran in 1979.
He received the B.Sc. degree in geomatics
engineering and the M.Sc. degree in geographic
information systems (GIS) from the K.N.Toosi
University of Technology, Iran, in 2002 and 2007,
respectively.
In 2006, he joined the Risk Management
Research Center, International Institute of
Earthquake Engineering and Seismology (IIEES),
Tehran, Iran as a GIS expert and a research
assistant. Since 2010, he has been a Ph.D. candidate in graduate school of
Informatics, Kyoto University, Japan.
His current research interests include GIS, multi-agent systems,
decentralized/distributed coordination, automated planning and scheduling,
emergency/crisis response systems, software agents, spatial agent-based
model, geoinformatics, and intelligent mobile GIS.
He was the recipient of the first prize in the GIS competition ranked first in
the national student competition, Iran, in 2001. He received the Japanese
Government Scholarship for the Ph.D. program in 2009. He was a visiting
scholar at both the Information Sciences Institute of University of Southern
California and the Robotics Institute of Carnegie Mellon University from
November 2011 to November 2012.
46

Dynamic assignment of geospatial-temporal macro tasks to agents under human strategic decisions for centralized scheduling in multi-agent systems

More Related Content

Viewers also liked (7)

Similar to Dynamic assignment of geospatial-temporal macro tasks to agents under human strategic decisions for centralized scheduling in multi-agent systems (20)

More from Reza Nourjou, Ph.D. (20)

Recently uploaded (20)

Dynamic assignment of geospatial-temporal macro tasks to agents under human strategic decisions for centralized scheduling in multi-agent systems