Mapping Cognitive Work Analysis (CWA) To An Intelligent Agents

Proceedings of the Defence Human Factors Special Interest Group (DHFSIG) 2002. DSTO Melbourne,
Australia, 21-22 November.
Mapping Cognitive Work Analysis (CWA) To An Intelligent
Agents Software Architecture: Command Agents
Frank Lui
Land Operations Division, DSTO, Edinburgh, Australia
Marcus Watson
ARC Key Centre for Human Factors & Applied Cognitive Psychology, Queensland University,
Queensland, Australia
Abstract: Currently, Land Operation Division is using computer generated forces to conduct operation
analysis studies. Intelligent Agents can potentially reduce the overhead on such experiments and studies
by reducing the amount of human resources needed. Command Agents, which are an external software
agent to the wargame simulation, can take the raw data, work out the context, plan how to carry out the
operation and assign tasks to subordinate agents. In addition, they receive information back from the
wargame — for example detection of enemy, fuel and ammunition status — and use this to build situation
awareness and to respond to unforeseen situations. Hence, these agents can take over many functions
performed by human operators in wargames. We are reporting on the process and the issues of using a
CWA study as a basis for modelling the decision making of a human operator.
1. Introduction
Land Operation Division, DSTO, Australia, have used Computer Generated Forces (CGF) to
simulate land battles in operation analysis studies for military exercises. Wargame simulation
program such as OneSAF Testbed Baseline (OTB) is used in the Synthetic Environment (SE) to
model the behaviours of constructive entities (i.e. Computer simulations for an individual
dismounted and mounted infantry, tanks, Armour Personnel Carriers (APCs), trucks, and aircraft of
any type etc.. Subsequent behaviours such as communication between entities, group assignments,
individual role and responsibilities in a platoon, company and battalion are also modelled.
In the SE, the human in the loop (HITL) system is a large part of the wargame. OTB, which is
Distributive Interactive System (DIS) compliant, can support HITL virtual simulations and human
operators (pucksters) to interact with the wargame thereby allowing them to perform operations in a
wargame as if they are in the real world. In a typical experiment, a puckster would be playing the
role of a battalion commander whose responsibility is to manage up to five companies. Practically,
the pucksters can be under a lot of pressure whilst fulfilling the battalion commander role as well as
issuing commands and moves via the OTB graphical user interface (GUI). Each puckster may
therefore be responsible for communicating with the Brigade Headquarter, planning and issuing
orders at the company and platoon level.
In an attempt to model the decision-making processes in land battle environments, Cognitive Work
Analysis (CWA)1 is used. CWA comprises of five steps namely; work domain analysis, control task
analysis, strategies analysis, social-organisational analysis and work competencies analysis. The
advantages in using this model is that it will provide us the tracability of decision-making in an
organisational structure and secondly, it provides the linkage between the abstract functions in the
higher hierarchy level and the plans and Courses of actions (COAs) performed in the lower
hierarchy level. However, just like other computational simulations, the Command Agent has a
limited selection of plans and COAs because of its rigid design. Because of this, the human players
will still have the final say in a situation. For instance, when the CA has exhausted all the possible
COAs and plans to be considered for a situation, it will send a request to the puckster for command
and control inputs as the result of a fail plan.
In this work, our aim is to use the Believe Desire Intention (BDI) based intelligent agent software
architecture, JACK-Teams2 and Command Agents (CA)3, to computationally model the decisionmaking processes made by a Company Commander.
The decision-making process involves
planning and making snap decisions of how to use troops based on situation awareness arising from
factors such as enemy position, enemy contacts, friendly troop morale, weapons status, fire support,
and terrain features etc. The challenge here is therefore to effectively model the behaviours of an
experienced commander who would normally take control of a situation by applying a combination
of best army practice and military doctrine4. Our objectives in this work are therefore: 1) to elicitate
the knowledge of expert domain such as an experienced Mechanised Infantry Company
Commander, 2) to document the knowledge in a systematic way and 3) to implement an algorithm
which encapsulates the thinking process, standard of procedures and the COAs required to execute a
plan. In this process, we will have mapped CWA to a BDI based software architecture such as the
Command Agents.
1.1 Simulation of a company commander using the command agent
In this work, we are concentrating on a company attack scenario in which the agent is to control a
motorised infantry company. The command agent is intended to carry out tasks now performed by a
human OTB operator. The role of the operator is to enter commands for the control units and entities
through a GUI. The OTB GUI comprises a map and icon display, a series of context-based editor
windows and a series of pull-down menus. The operator and commander can use the graphical
representation of the wargame on the GUI to acquire an adequate level of situation awareness,
undertake terrain appreciation and monitor the progress of the battle.
In current use, objectives given to the OTB operator who follows military doctrine and standard
procedures to produce plans for courses of actions to be taken during the battle. With command
agents, the objectives of the attack are inputs to the agents. The agents will then use their reasoning
and rules, which are based on military doctrine, to produce plans and courses of action. Following
this, the tasks and commands are sent to the simulation in a format that is understood by the
simulation program. During execution of the task, the agent monitors progress and makes any
necessary adjustments to the plan, thereby fulfilling the supervision and control function of the OTB
operator.
1.2 Concept demonstration of the command agent software
To demonstrate the functionality and capability of the command agent, we have chosen a deliberate
attack scenario (see Figure 1). The objective is for a
mounted infantry company located at Point A to attack
an enemy formation occupying a position in the
A
vicinity of Point B. The company commander agent is
to produce a plan and courses of action to carry out
four phases for the attack, namely 1) preparatory, 2)
assault, 3) exploitation and 4) reorganisation4.
The company organisation consists of three platoons.
Each platoon is comprised of three sections, each of
which has nine soldiers and an armoured personnel
carrier (APC). To prosecute the attack, the company
splits into a fire support platoon and two assaulting
platoons. The agent plans the routes, form-up positions
and coordination parameters for the attack. The agent
will monitor the location and status of its own troops
and the enemy and will respond to situations that
require changes to the plan. In planning and executing
the attack,
the agent will apply documented military
4
doctrine and make appropriate use of terrain.
2.
Fire support
B
Figure 1: The scenario for demonstrating the
agent capabilities. The objective is for a
mounted infantry company located at Point
A to attack an enemy formation occupying a
position in the vicinity of Point B.
Military Knowledge Elicitation
Developing a command agent to undertake the required tasks necessitates an adequate knowledge of
military doctrine in a form that can be built as rules for the operation of the agent. High fidelity
knowledge elicitation methods5 can be used to extract rules and reasoning from specific military
expert domains. This was achieved by using a Cognitive Work Analysis (CWA) approach. This
process requires several interviews with experienced company commanders. During the interviews,
they are presented with questionnaires that are designed to elicit knowledge of how decisions are
made during combat.
However, one must bear in mind that wargames, such as OTB, have limited fidelity. For instance, in
real life, when an infantry unit is performing a reconnaissance mission, its tasks are to patrol a
specified region and to report on the enemy’s status and position. The infantry unit tasks include
report enemy detections, avoid line-of-sight (LOS) with its opposition and report the type of enemy
weapons used, how they are used, the enemy’s strength and activities etc. The command agent’s
tasks are, therefore, 1) to move the infantry unit to the destination, 2) check for LOS with enemy
forces, 3) report on enemy detection, 4) collate the information and put them in context relevant to
the wargame scenario and so on. What is missing here is the impact that the enemy’s activities and
doctrine have on the company commander’s decisions. In wargame simulations, this is classified as
the enemy’s beliefs and intent. Very often, this information is very hard to determine using the rules
that are built into the agents. In order to achieve a consistency and coherence in information flow
throughout the simulation, it is necessary to extract only those rules that are relevant to the context in
the wargame scenario, from the CWA model.
3. Suggestions of how to apply the CWA to CA development
The five phases of the CWA all contribute to developing a template for Company commander
decision-making. Work Domain Analysis (WDA) can be used to elicit relationships between the
three different WDA namely; Mechanised Infantry, Environment and Enemy. This is followed by a
Control Tasks Analysis that is in the form of Decision Ladders (DL) and can beordered into
Temporal Coordination and Control Tasks (TCCT), to plan and execute Company attacks. Finally,
Worker Competency (WC) and its effect on the selection of COA are also included in the model.
3.1 Integrating WDA Objects into the SE
This work has split the Physical Objects of the WDA the in four distinct groups: Formation Objects,
Communication Objects which represent real Physical Objects; Supply Objects which represent
consumable Physical Objects; and Map Objects which are imposed limitations of the commander’s
planing. Supply Objects will have representations within the SE. Some Supply Objects might,
however, only be modelled in terms of their ability to hit and destroy an enemy target rather than the
Beaten Zone or Impact Area of the weapons; this may be problematic for COA planning.
Commanders take into consideration the ability of weapons to neutralise enemy forces as well as to
destroy them. Therefore a commander can assess the ability of a weapon or formation to neutralise
Beaten Zone / Impact
Safe limit for vehicle
infantry under cover
Safe limit for infantry
in the open
81mm
Figure 2: Beaten Zone and impact overlay templates
an enemy position that can be covered in the
Beaten Zone or Impact Area. Figure 2 illustrates
the factors a commander must consider with such
Supply Objects.
Possible direct
fire support
position
B
FUP
Company
boundary
Direct fire must
cease as the
when the
assault force
reaches the
Templates can be utilised for other aspects of the
WDA, such as Map Objects —FUP (Form Up
Places), boundaries etc. Interactions with the
Axis of
advance
templates can then be used to assess the feasibility
of each COA. For example, the SE will represent
A
effective weapon ranges, which can be compared
against a template to determine if an approach
Possible direct
fire support
would risk fratricide. In Figure 3, Position A
position
Company
boundary
indicates a direct fire support position that offers
visual and direct fire affordance (no topographical
obstacles between A and the target), the maximum
time that fire can be brought to bear on the target,
Figure 3: Templates to determine boundary
and protection. If a direct fire support position or
limitations on COA selection
assault approach would project friendly fire
(Position A) across boundary templates into another unit’s area of operation, then such a COA would
be unfavourable. To use this position would require permission from the Battalion commander of the
unit across the boundary.
Position B offers less time for fire from this position to be on the target, as this will have to lift
earlier than Position A, because of the safety limit of the beaten zone. It does not, however, project
fire into adjacent unit areas of operation and therefore may be the better position. If the position can
be reached without enemy observation (enemy visible template) and within the time frame of attack,
then it becomes a viable direct fire position. Even if the SE does not model the effect of munitions
that miss their target, such comparisons of weapon ranges compared to company boundaries can be
useful in selecting approaches and direct fire support positions that could be selected by a human
commander in a real conflict.
3.2 Modelling CA decision-making competency
We have identified three methods in which Worker Competency (WC) can be incorporated into the
SE. First, WC could be modelled at the domain knowledge level, where different CAs allocate
different weightings to different relationships within the set time. Second, WC could be reflected in
the order different DLs are evaluated. The third way is to simply assign a global accuracy for
evaluating COA. CAs representing less experienced commanders and support staff could have a
greater discrepancy applied to the combat-factor for each COA. The global accuracy could also be
applied to the assessment of future tasks the company is likely to receive in relation to the selection
of COA.
The modelling of WC for CAs is less important than the control tasks and the domain. WC does
offer insight into the difference between diverse military backgrounds, which may force
commanders to consider the constraints and affordances of subordinates as well as their formations.
The inclusion of differences between CA due to WC will also need to be reflected in the information
provided and feedback given to human commanders wargaming in the SE as commanders normally
would have some understanding of their subordinates’ abilities.
4. Current and future work
Currently, we are in the process of developing the algorithm for the 4 phases in a deliberate attack
namely, Preparatory, Assault, Exploitation and Reorganisation. To implement reasoning for terrain
and firepower, templates of terrain features and weapon fire range is created and integrated into the
command agent to determine routes to potential targets and objectives. From this, the command
agent can assign tasks (such as move vehicles, attack a target and occupy area etc) to the motorised
infantry units. Reporting capability is generated when certain events such as enemy contact, fire or
hostile aircraft detections are triggered. This feature will give the command agents the capability of
aggregating the raw data which is sent by OTB, and to generate reports for both human operators
and other agents in the simulation environment.
To implement CWA model in the command agent software, we have conducted interviews with
various army personnel to elicitate military knowledge on the use of fire support team and the tactics
of maneuvering assault teams in an attack. This enabled us to define the relationships between
different levels of abstract functions in the WDA. By incorporating CWA in the command agent,
the command agents will hopefully have the capabilities of selecting plans based on the weighting of
certain COAs which are dependent on the conditions of detected enemies, their fire power, friendly
troops and terrain features.
5. Conclusion
In conclusion, we are currently integrating all the required components for the command agent
namely; JACK-Teams, message aggregation, OTB and terrain appreciation. We will be
incorporating the CWA model to assist with modelling the commander of a company size
mechanised infantry which will be able to take control of its platoon unites and execute higher orders
given by battalion commander. With this capability, the command agent will be able to perform
some of the routine functions that are normally done by a human puckster in a war game exercise.
This will greatly reduce the overhead cost in conducting war game exercises and experiments.
6. Acknowledgements
We would also like to thank Command Agents software team members in Land Operations Division,
Edinburgh, Australia and Dennis Jarvis and Jacquie Jarvis from AOS Pty Ltd, who are currently
participating in the implementation of the command agents and Teams software.
7. References
1
Penelope M. Sanderson and Marcus Watson, “Cognitive work analysis and analysis design and
evaluation of human-computer interactive systems”, OZCHI 98, designing the future, conference
proceedings, 30th Nov to 4th Dec 1998, Adelaide, South Australia, p220-227.
2
Lucas, A., Rönnquist, R., Howden, Hodgson, A., Connell, R., White, G. and Vaughan, J.,
“Towards Complex Team Behaviour in Multi-Agent Systems”, Proc. SimTecT 2001.
3
Frank Lui, Jon Vaughan, Russell Connell, Dennis Jarvis and Jacquie Jarvis, “An Architecture to
Support Autonomous Command Agents for OneSAF Testbed Simulations”, Proceeding of SimTect
Conference, 13-16 May 2002, p275-280.
4
Manual of land warfare part two infantry training, Vol. 1, Pamphlet 2, The Rifle Platoon, 1986, (MLW
2-1-2).
5
Dan Diaper, “Knowledge elicitation principles, techniques and applications”, Ellis Horwood Limited,
Publishers, Chichester, 1989.