RESEARCH AND DEVELOPMENT
Using Graphics Processor Units to
Accelerate OneSAF: A Case Study in
Technology Transition
Massively Multi-Player (MMP)
Environments for Asymmetric Warfare
Simulating Believable,
Context-aware and Culture-specific Human Behaviors
A Graph-based Approach for
Automatic Building Extraction from Aerial LIDAR Data
Automatically Converting Paper Map
Images For Use As Notional Terrain
An Ontological Data Structure for
Real-Time Simulation
A DSAP Infrastructure for
Analysis, Training, and Operational Support
Empirical Foundations for
Intelligent Coaching Systems
Automatic Detection of
Discrepancies in After Action Review
Accurately Representing Target
Distance in a Flight Simulator
Automating Simulation-Based Air
Traffic Control
Depth of Focus and Perceived
Blurring of Simultaneously-Viewed Displays
Towards Improving the
Instructional Design Process for Team Training
Developing Situation Awareness
Metrics in a Synthetic Battlespace Environment
Simulation Data Grid: Joint
Experimentation Data Management and Analysis
Customizing Interactive Training
Through Individualized Content and Increased Engagement
Artificial Intelligence for
Constructing Accurate, Low-Cost Models and Simulations
Simulating Ethnic Conflict and
Secessionism for Joint Experimentation
Using Graphics Processor Units
to Accelerate OneSAF: A Case Study in
Technology Transition
Marlo Verdesca, Jaeson Munro, Michael Hoffman Science Applications International Corporation Maria Bauer RDECOM Dinesh Manocha
On-going research aims to accelerate the runtime processing speed of the One Semi-Automated Forces (OneSAF) Computer Generated Forces (CGF) simulation by converting and migrating some of the core algorithms from the host Central Processing Unit (CPU) to an on-board auxiliary Graphics Processor Unit (GPU). In this research the GPU chip is regarded as a surrogate stream processor and appropriate algorithms are designed to map to the GPU architecture. Processing speed gains are realized both through computational capabilities of the GPU as well as through offloading of the host CPU. Technology transfer of this research into the OneSAF user baseline is a key requirement of this research. The OneSAF development program
focuses on the same issues of scalability and runtime performance that will
be directly affected by use of GPUs. As program architects are marshalling
conventional approaches for resolving these challenges, the introduction of
GPU-based solutions is being realized.
This paper examines the challenges, planned approaches and benchmarked
results for using GPUs to accelerate OneSAF simulation. 2005 Paper No. 2121 |
|
|
Massively Multi-Player (MMP)
Environments for Asymmetric Warfare
Michelle Mayo Dr. Michael J. Singer, PhD U.S. Army Research Institute Laura Kusumoto Forterra Systems, Inc. It has been proposed that the Army needs a high-level training capability for asymmetric missions. Current training programs are focused on conventional warfare and are mainly limited to units that are co-located. Where training capability exists, the scenarios have limited interactivity and fail to address a variety of cross-cultural communication issues that troops encounter in the real world. U.S. Army's Research, Development and Engineering Command,
Simulation and The paper details research in the formative evaluation of internet-based training using Soldier participants and gaming technologies. Initial test results with the 101st Airborne Division of Fort Campbell, Kentucky (KY) and the Illinois Army National Guard are presented to indicate the potential such technology has to meet new asymmetric training needs and optimize use of Soldiers’ time while preparing for live training events and actual deployment. The paper also addresses the tools needed to build the training environments and required After Action Review capabilities. 2005 Paper No. 2149 |
|
|
Simulating Believable,
Context-aware and Culture-specific Human Behaviors
Edward M. Sims, Ph.D. Vcom3D, Inc.Orlando, FL Successful peacekeeping missions require the ability to identify civilians’ needs and intentions, and to successfully influence or direct their actions. For these missions, interpreters perform an important role in mediating our soldiers' interactions with the local population. However, research has shown that up to 93% of human communication comes through non-verbal, primarily visual channels. In many ways, understanding nonverbal language and unwritten rules of human interaction for a region can be even more important than spoken language skills. In the case of nonverbal communication, we often see behaviors that we think we understand, or use behaviors that we think are understood, but which really have a very different meaning in another culture. These misunderstandings can make the difference between successful and failed missions. This paper describes the challenges in simulating virtual actors with the capability to act on high level "stage directions" in a way that considers the spatial, temporal, and social context of a situation. Specifically, we will review research that demonstrates the need to accurately portray focus of attention, gesture, facial expression, and “body language”. We will describe a Behavior Engine, an associated Behavior Knowledgebase, and a methodology for creating and extending this knowledgebase. Using this technology, we will demonstrate the importance of context and culture in behavioral simulations by evaluating the effects of incorporating the proposed techniques. Finally, we will demonstrate how we are currently using this technology to enhance Interactive Multimedia Instruction, and how we will adapt the technology for DARWARS game-based instruction. 2005 Paper No. 2291 |
|
|
A Graph-based Approach for
|
|
|
Automatically Converting Paper
Map Images For Use As Notional Terrain
William T. Richards VMASC / ODU David Hibler PhD. This paper will detail the results of a thesis project
conducted to explore the problems and automated solutions for preparing a
real world map image similar to Archival Research Catalog Digitized Raster
Graphics (ADRG), for use as a notional terrain map image during a simulation
exercise. One of the problems with
notional terrain occurs when the original map image is stitched in to the
exercise event terrain, the paper map image may no
longer accurately represent the terrain.
This is especially true if there has been a change in the latitude,
orientation, or size of the original image.
Additionally, other problems can arise if the names, labels and route
numbers of the features on the map are changed for the scenario. The research conducted shows how it is
possible to automatically identify and remove undesirable features from a paper
map image and how to automatically restore some of the features that would
have been underlying the removed objects.
This is done in an effort to retain the amount and type of naturally
existing errors between the paper map image, vector data, and imagery data,
and also allow the cartographer to easily place new names and labels on the
cleaned map to match the scenario with a minimum of effort, while adding to
the overall experience of the training audience. 2005 Paper No. 2052 |
|
|
An Ontological Data Structure
for Real-Time Simulation
Barry Bitters Department of Environmental Studies In recent years, feature classification systems have proven to be inadequate in their ability to assist in the detailed description of synthetic cultural and natural environments. Many existing feature classification systems were developed when limited display capabilities demanded economy in the database generation process. Others are so limited in scope that they do not provide the feature specificity necessary for current and future display capabilities. Based on current client requirements and the anticipated detail requirements of future synthetic environment databases, an alternate taxonomic structure must be available to database developers. In a presentation at I/ITSEC-2002, we described a conceptual feature classification system, designed for use in the development of synthetic environments. We described the potential uses and a conceptual design of a hierarchical data structure that could be used to store natural and cultural feature data. The first version of this taxonomy is now available for
public review and possible implementation in existing and future software
systems. The development of this new system was based on creating a logical
intersection of classes and concepts from many existing taxonomies, thesauri,
ontologies and classification schema. This new data
structure now allows unambiguous inventorying of natural and cultural
landscapes. It can also assist in the unambiguous importing of disparate
sources of feature data. A significant portion of this research/production
effort also involved the creation of a public-domain library of generic 3D
models – models closely aligned to the concepts in the feature taxonomy.
These data structures and their availability are discussed in detail. The
trials and tribulations encountered during development, potential uses of the
data structure, and future developments to be performed under this research
effort, are also discussed. 2005 Paper No. 2212 |
|
|
A DSAP Infrastructure for
Analysis, Training, and Operational Support
Craig Lammers, Robert McGraw RAM Laboratories, Inc. Dawn Trevisani Air Force Research Laboratory, IFSB RAM Laboratories and the Air Force Research Laboratory’s Information Directorate are developing a software infrastructure to provide a Dynamic Situation Assessment and Prediction (DSAP) capability for analysis, training, and operations through current state estimation and prediction capabilities. These capabilities will allow Commanders to evaluate and analyze Courses of Action (COAs) and potential alternatives through real-time and faster-than-real-time simulation by executing multiple instances of the Joint SemiAutomated Forces (JSAF) simulation (or a simulation of choice) simultaneously across a computing grid. The infrastructure supports analysis by executing and evaluating simulations in an analytical capacity as a form of predictive mission rehearsal. The infrastructure supports training through the use of the real-time simulation element that supports human-in-the-loop that allows trainees to see the result of their actions, while also providing feedback to some of the alternative actions that could have been selected. The operations aspect to the infrastructure allows for tracking the internal state of the operational environment via the real-time simulation, predicting outcomes of plans and alternatives through faster-than-real-time simulation, and calibrating both the state estimation and predictive elements of the system with Red and Blue Force data extracted from real-time data feeds, which can be used to track the re-tasking of assets in support of Time Sensitive and Time Critical Targeting. This paper covers the current state of the art for DSAP
while examining the steps necessary to transform the DSAP infrastructure to a
Global Information Grid (GIG)-ready tool. This work includes (1) defining the
granularity to which elements of the DSAP infrastructure will be broken down
for use in a GIG environment; (2) defining the types and format of data and
messages that will be transmitted between elements of the infrastructure and
resources residing on the GIG, and (3) defining the web services that will be
provided by DSAP. 2005 Paper No. 2152 |
|
|
Empirical Foundations for
Intelligent Coaching Systems
Elizabeth Owen Bratt, Karl Schultz, Stanley Peters, Tina Chen Heather Pon-Barry Robert Bosch Corp. The Navy is shifting its training and education from traditional methods, such as on-site instruction, texts, and observing students during drills, to computer-supported learning such as web-based instruction and computer simulations in lieu of live drills. This transition presents the challenge of keeping the best parts of traditional methods of instruction while obtaining the advantages that computers afford. The challenge is more difficult because to maximize savings in manpower, money and time, computer-based learning must be able to teach, evaluate and give feedback to students without any instructor in the loop. A valuable aspect of traditional training methods, in which computers currently fall short, is the mentor/student relationship: an experienced person monitoring and guiding a novice's performance. The mentor gives the student direct, personalized feedback in a setting where the student can ask questions and discuss issues. Most computer simulations are lacking in this type of interaction. We propose that giving computers the ability to hint,
question, prompt and guide a student's actions using natural language will
more closely simulate this relationship and greatly improve the effectiveness
of computer-based learning. To assess this hypothesis, we are utilizing
natural language technology to (1) allow students to use a damage control
trainer for surface ships by speaking with the simulation system, and to (2)
support a concurrent spoken discussion with an intelligent coaching system
that aims to improve the student's immediate and future performance. The
combined system performs a mentoring function, helping students learn correct
actions and avoid practicing mistakes. We present data from 2005 Paper No. 2389 |
|
|
Automatic Detection of
Discrepancies in After Action Review
Joakim Ekblad, Avelino Gonzalez Hans Fernlund Paul Barath Saab Training Systems AB After-Action Review (AAR) is an effective tool to evaluate and improve the performance of trainees in tactical training exercises. However, when the exercises grow in size, and might reside in several locations, providing feedback to the majority of the participants can be complicated. It requires extensive time and resources, and the review might be limited to the few most important tactical decisions made. This paper presents a model of how to automate the After-Action Review and make it easily accessible to all the participants to increase the efficiency and improve the performance of After-Action Reviews. A system built on expert models where the action of the trainees could be compared with these models can provide additional support for the trainees. However, such a system needs to automatically detect and classify discrepancies. Discrepancies between a trainee and an expert modeled agent can emerge in many situations. By minimizing the discrepancies shown in the AAR to only include the ones believed to be significant enough to decrease the performance of the trainee, the AAR will become more effective by reaching out to the majority of the participants of the exercise giving them individual performance feedback. Preliminary results of our experiments are promising and indicate that the model presented in this paper can be used to address the issues discussed above. 2005 Paper No. 2404 |
|
|
Accurately Representing Target
Distance in a Flight Simulator
George A. Geri Link Simulation and Training Christine M. Covas, Marc D. Winterbottom, & Byron J. Pierce Air Force Research Laboratory Rendering an essentially continuous IG image onto the
discrete-pixel array of most display devices, as well as graphical processing
such as antialias filtering, can result in
significant variations in displayed target size and hence simulated
distance. Using videotape recordings
obtained from a CRT-based, flight-simulator display, we have directly
measured changes in size, over a three-second simulation interval, of target
aircraft simulated at distances between 3000 and 11000 feet. In Experiment 1, the percentile of the
measured target-size distribution which corresponded to the nominal target
size was found to change with simulated distance. Additionally, an interaction was found
between pixel count (1280×1024 and 2048×1536) and antialias
filtering (0 and 2×). In Experiment 2,
a single intermediate pixel count (1600×1200) was tested, and in addition,
eight target gray-levels were tested perceptually, in order to directly
compare the videotaped imagery with what was visible on the screen. It was found that the percentile
corresponding to nominal target size varied with both simulated distance and antialiasing condition (0 and 4×). In both experiments,
for the larger (i.e., closer) targets, the nominal target size corresponded
to about the 96th percentile of the distribution of measured sizes. As target size was decreased, nominal size
was found to correspond to as low as the 60th percentile. Further, in both experiments, the functions
relating the relevant size-distribution percentiles to simulated distance
were nonlinear, and in Experiment 2 they were different for each of the antialiasing conditions tested. The data indicate that the average size of
targets displayed in CRT-based flight simulators is smaller than would be
expected from their nominal distance as defined by the IG. In addition, the unexpected complexities found
in the size-distribution data indicate that accurately adjusting displayed
target size to reflect a chosen target distance will require corrections that
are dependent on simulated distance. 2005 Paper No. 2147 |
|
|
Automating Simulation-Based Air Traffic Control
Glenn Taylor Soar Technology, Inc. MSG Jeff Miller (US Army, Ret.) Navigator Development Group Inc. Jeff Maddox Air Traffic Control (ATC)
receives little attention in simulation-based training and experimentation,
in part because of the cost of including human operators to play ATC roles.
Where ATC is used, it is typically very limited, reducing the realism of the
experiment or training experience. This problem has become
more apparent as UAVs and as joint battles are
more often fought in simulation, requiring closer human management of the
simulated airspace to coordinate air corridors, restricted airspaces, joint fire support, and the like. Furthermore, UAVs have become more prevalent in real battlefields, and the services are struggling with
how to employ them safely and effectively within a broader air operations
picture. Fighting ATC realistically in a simulated battlespace
can help develop more realistic and appropriate employment tactics in the
real battlespace. This paper describes the results
of a Phase I SBIR investigating the feasibility of automating air traffic control (ATC) within simulation
environments, for both experimentation and training. We leverage prior
research analyzing human ATC tasks and situational awareness requirements in
Tower, TRACON, and En Route operations, and describe how simulation
environments can place different constraints and requirements on an ATC
capability. We describe the use of human-driven ATC in recent joint
experiments as a way to define some
operational requirements of automated ATC.
Key requirements include the ability to interact with both human
pilots in virtual cockpits (using voice interaction), and with synthetic
pilots and existing airspace management tools (using digital data links). We
identify existing tools and technologies that can be used to fill these requirements, and where technology gaps
still exist. Finally, we describe a cognitive systems approach to automating
simulation-based ATC, and the development of a limited prototype that
illustrates some of the key components of the architecture. 2005 Paper No. 2193 |
|
|
Depth of Focus and Perceived
Blurring of Simultaneously-Viewed Displays
Marc D. Winterbottom, Byron J. Pierce, Christine Covas Air Force Research Laboratory Robert Patterson Jennifer Winner Link Simulation and Training Head-mounted displays (HMDs) have
not previously been combined with flat-panel display systems and it was
unknown whether viewing two displays at differing focal plane distances would
lead to perceived blurring or visual discomfort. This is now a concern as the Joint Helmet
Mounted Cueing System (JHMCS) is integrated with existing flat-panel display
systems such as the Mobile Modular Display for Advanced Research and Training
(M2DART). The degree of blurring that
could occur would be dependent upon observers’ depth of focus and the extent
to which the two displays vary in focal plane distance. In previous research, we investigated
whether blurring occurs when two displays are viewed simultaneously at
independently varying focal plane distances.
These conditions simulated those of a monocular HMD integrated with
the M2DART. The results of that
research suggested that blurring due to two differing focal planes was not
likely to be a significant issue for the current configuration of the
M2DART. We present here two additional
experiments that extend these earlier results. In the first experiment, luminance levels
were decreased, thus increasing pupil size and decreasing depth of focus and
the degree of blurring was measured using psychophysical techniques. In the second experiment, blurring and
visual discomfort were examined under more typical viewing conditions: observers performed a task similar to
off-bore sight targeting in the M2DART using a monocular HMD. They identified the orientation of an
aircraft target presented on the M2DART and a test letter presented on the
HMD. Assessments of eyestrain and
perceived blur were obtained during the performance of this task. The results of these two experiments
indicated that depth of focus should not be an issue for standard-resolution
displays and, further, that visual discomfort is not likely to be an issue
for the integration of a monocular HMD with the M2DART. 2005 Paper No. 2273 |
|
|
Towards Improving the
Instructional Design Process for Team Training
Marcel van Berlo TNO Defence, Security and Safety Business Unit Human Factors Soesterberg, The Netherland Conducting team training is daily business for the military. Designing team training programs and exercises, however, is not always that structured. Instructional designers are in fact trained to design instruction primarily for individuals. After their instructional design course, they learn to design team training more or less on the job. A literature and field study show that instructional designers for team training find it hard to conduct task analyses at the level of a team: it is difficult to capture the interactivity and interdependency between the various team members, and to formulate team instructional objectives. Without this proper input, designing team training scenarios is difficult: paying attention to all relevant team task and team work aspects in a systematic way does not occur automatically. This process may be improved by offering these instructional designers adequate support. During three design-experiments, we developed and tested guidelines and a workshop supporting the analysis of team tasks and the design of team training scenarios. For the first (task analysis) and second (scenario design)experiment, two versions of guidelines were developed: an experimental version with an explicit focus on team aspects, and a control version in which this specific focus was absent, resembling traditional guidelines. The results of the first design-experiment show that the experimental guidelines lead to a significantly better quality of the analysis process; the results of the second design-experiment show no significant effects. The purpose of the third design-experiment was to investigate the effect of a more elaborate introduction (an interactive workshop) of both sets of experimental guidelines. The results show that only on topics that were explicitly dealt with during the workshop, the quality of the analysis and design process improved. 2005 Paper No. 2037 |
|
|
Developing Situation Awareness
Metrics in a Synthetic Battlespace Environment
Jacqueline M. Curiel, Michael D. Anhalt Alion Science and Technology John J. Tran, Ke-Thia Information Sciences Institute/USC The Joint Forces Command (JFCOM) conducts Joint Urban
Operation (JUO) exercises in synthetic battlespace
using human-directed computer simulation tools such as Joint Semi-Automated
Forces (JSAF) to support ongoing joint war-fighting efforts. A component of these experiments is that of
human-in-the-loop (HITL) interactions where human players impact the outcome
of the exercise. This is in contrast
to Building upon work presented at I/ITSEC 2004 (An Interdisciplinary Approach to the Study of Battlefield Simulation Systems, paper 1886), we adopt a cognitive-computational approach for measuring SA based on Situation Model theory. Situation models are complex mental representation of events. As events unfold, these mental representations must be updated to maintain an accurate representation. Prior research has demonstrated that situation models are updated along a number of dimensions. These dimensions reflect information about entities, space and time coordinates, participants’ goals, and the causal relationships of events. We utilize the information encapsulated in SA objects (SAOs), recorded during the JUO exercises, to develop a tool that automatically monitors players’ SA and evaluate the importance of these dimensions on situation awareness over the time course of the experiment and on the three levels of SA. Our findings have practical implications for subsequent training, product development, and extend the knowledge base of cognitive behavior. 2005 Paper No. 2218 |
|
|
Simulation Data Grid: Joint
Experimentation Data Management and Analysis
Ke-Thia Information Sciences Institute The need to present quantifiable results from simulations
to support transformational findings is driving the creation of very large
and geographically dispersed data collections. The Joint Experimentation
Directorate (J9) of United States Joint Forces Command (USJFCOM) and the
Joint Advanced Warfighting Project (JAWP) is
conducting a series of Urban Resolve experiments to investigate concepts for
applying future technologies to joint urban warfare. The recently concluded
phase I of the experiment utilized and integrated multiple scalable parallel
processors (SPP) sites distributed across the Building on lessons learned in developing data management tools for Urban Resolve, we present our next generation data management and analysis tool, called Simulation Data Grid (SDG). The design principles driving the design of SDG are 1) minimize network communication overhead (especially across SPPs) by storing data near the point of generation and only selectively propagating the data as needed, and 2) maximize the use of SPP computational resources and storage by distributing analyses across SPP sites to reduce, filter and aggregate. Our key implementation principle is to leverage existing open standards and infrastructure from Grid Computing. We show how our services interface and build on top of Open Grid Services Architecture standard and existing toolkits (Globus). SDG services include distributed data query/analysis, data cataloging, and data gathering/slicing/distribution. We envision SDG to be a general-purpose tool useful for a range of simulation domains. 2005 Paper No. 2292 |
|
|
Customizing Interactive Training
Through Individualized Content and Increased Engagement
Brian Magerko Robert E. Wray, Lisa
Scott Holt, Brian Stensrud Soar Technology,
Inc. Simulation-based training offers the potential for
relatively low-cost training available at any time and almost any duty
station. However, a main drawback of simulation-based training is the lack of
oversight in the training process. Simulations often depend on a fixed number
of pre-defined training scenarios that are designed to test training
objectives but not to deliver a training experience customized to the
specific trainee’s current level of skill and understanding. In this paper, we introduce the Interactive
Storytelling Architecture for Training (ISAT), which uses an intelligent
agent, the director, to assemble training scenarios that test the skill level
of individual trainees. The director also provides indirect feedback about
trainee actions during the execution of a training scenario, subtly adapting
the training environment to stress unmastered
skills and suggest remediation. This
approach results in a training experience that is specialized to the
trainee’s individual needs and potentially more engaging, resulting in faster
development of trainee proficiency. 2005 Paper No. 2000 |
|
|
Artificial Intelligence for
Constructing Accurate, Low-Cost Models and Simulations
Dr. David P. Brown CAPT (ret) Richard A. Mohler, USN Northrop Grumman Corp. Modeling and Simulation is an important tool in the
development of the highly effective weapons systems built by the 2005 Paper No. 2009 |
|
|
Simulating Ethnic Conflict and
Secessionism for Joint Experimentation
Alok Chaturvedi R. Chaturvedi, C.M. Foong, M. Mulpuri, D. Lengacher, S. Mellema Simulex, Inc The post Cold War period has been characterized by a dramatic increase in intra-state conflicts, the later significantly outpacing inter-state wars both in their duration and intensity. In an increasingly globalizing and interdependent world intra-nation conflict in the guise of ethnic conflict has serious implications for regional and international stability. There is a burgeoning literature on the relative salience of political, economic, and social conditions in the emergence, escalation, and diffusion of ethnic unrest. Contributing theories analyze the significance and/or interaction of conditions such as deprivation, inter-group antipathy, institutional constraints, demographic change, predatory groups, and external intervention on ethnic awareness, mobilization, unrest, and subsequent irredentist goals. There is also a plethora of research on strategies for managing intra-state conflict. While conspicuous attempts are being made to synthesize explanations and prevention of ethnic unrest and secessionism, few methodological tools are available that can fully integrate the theories and strategies at various levels of a socio-political system—individual, group, national, and international. This paper presents a multi-agent simulation as a technique to explore, test, and validate theories of ethnic conflict at multiple levels of analysis. Specifically, it presents the development and implementation of a ‘Virtual State (VS)’ and the subsequent ‘Virtual International System (VIS)’ to explore: conditions for the emergence of ethnic unrest, conditions for the diffusion and escalation of ethnic conflict, conditions for the emergence and success of secessionist movement, and conditions for the success of
multilateral interventions and the effects of DIME (Diplomatic, Information,
Military, Economic) actions on
prevention of secessionist movement. 2005 Paper No. 2354 |
|
|
Training for Multicell and
Dismounted Command and Control
Paula J. Durlach U.S. Army Research Institute for the Behavioral and Social Sciences The Multicell and Dismounted
Command and Control (M&DC2) program is a Defense Advanced Research
Projects Agency (DARPA) initiative to
explore a new approach to Battle Command and Control using distributed
networked manned and unmanned systems.
The M&D C2 program continues the work started in the DARPA Future
Combat Systems Command and Control (FCSC2) program. The program modeled a network-centric
Command and Control (C2) prototype that supported execution-based operations
with distributed manned and unmanned entities. To date, the prototype has gone through six
iterations and has been used to conduct seven discovery experiments. As the iterations progressed, the systems
became more complex. In addition, novice operators began participating in
experiments, and it became clear that training would become a major
challenge. This paper focuses on the training challenges and discusses the
potential of the DARPA testbed to expand into the
realm of training research. Such
research could help establish the optimal training methods for changing a
Current Force Soldier into a Future Force Soldier. 2005 Paper No. 2063 |
|
|
Evolving Human Behavior Models
from Live Exercise Data
Hans Fernlund Avelino Gonzalez, Joakim Ekblad Paul Barath Saab Training Systems AB It is widely accepted that the difficulty and expense
involved in acquiring the knowledge behind tactical behaviors has been one
limiting factor in the development of simulated agents representing
adversaries and teammates in military and game simulations. Several
researchers have addressed this problem with varying degrees of success. The
problem mostly lies in the fact that tactical knowledge is difficult to
elicit and represent through interactive sessions between the model developer
and the subject matter expert. This paper describes a machine learning
approach to evolve tactical agents based upon automatic observation of a
human performing a mission. The approach employs Genetic Programming in
conjunction with Context-based Reasoning and is called Genetic Context
Learning (GenCL). Prior research collected data
from a simulator and built a prototype that showed the feasibility of the
approach. Conversely, this research collected data from live exercises when
two tank platoons were engage in a combat. The GenCL
algorithm learned the behavior of the tanks and incorporated the evolved
models into a simulator, aimed to be used in a support system for exercise
evaluations. When the data collection task transfer from simulated environment
to real world problems the degree of freedom dramatically increases. It opens
new questions and future research areas. However, two major contributions
were accomplished with this work: 1) The GenCL
algorithm is able to automatically create human behavior models from
pre-processed data collected from real exercises. 2) The results indicate
that humans inhibit behavior patterns that can be modeled in a contextual
manner. 2005 Paper No. 2300 |
|
|