AIR DEFENCE OPERATIONAL ANALYSIS USING THE SWARMM MODEL

David McIlroy, Bradley Smith, Clinton Heinze and Michael Turner
Air Operations Division, Aeronautical and Maritime Research Laboratory, DSTO,Department of DefenceMelbourne, Victoria, Australia
david.mcilroy@dsto.defence.gov.au

Abstract

Research and analysis of military operations using mathematical modelling requires a diverse blend of technologies ranging from vehicle performance and physical environment modelling, software engineering, simulation techniques and OR techniques, all of which must be tailored to achieve the required research outcomes.

Additionally military operational analysis requires researchers to have doctrinal and tactical knowledge of military operations in order to represent the tactical decision-making in assessing overall weapon systems performance.

This paper describes the application of SWARMM to analysis of complex tactical behaviour, in the domain of fighter air defence. SWARMM includes tactical reasoning software based on artificial intelligence, together with a convincing representation of physical systems. It is a powerful tool and research vehicle for the study of modern weapon system performance, including critical human decision-making.

The SWARMM model is being employed by AOD (DSTO, Department of Defence) in Defence studies of optimum fighter tactics, including evaluation of possible systems upgrades. SWARMM is now the preferred model in Air Operations Division for operational analysis studies from engagement to mission level. The paper will present a brief overview of SWARMM, the analytical techniques used and issues addressed, and some representative examples illustrating the advantages offered to the analyst in terms of speed and credibility for Defence customers.

1. SWARMM - A BREAK FROM TRADITION

In the branch of operational analysis practiced for military studies, the traditional approach has been to construct or employ existing mathematical models representing the behavioural and physical performance of the vehicles and their environment.

When studies are significantly influenced by the operational and tactical behaviour of the vehicle crews, it is clearly important to model these aspects with some fidelity and clarity. Hindsight shows that the modelling of the military doctrine and tactics presents significant challenges. These challenges arise due to the software design difficulties of modelling the tactics adequately and the associated development and testing timescales.

Recognition of the barriers to developing models and conducting studies efficiently led to a new approach employing software-agents featuring the Beliefs-Desires-Intentions (BDI) paradigm. An example of the use of this new paradigm is the SWARMM model.

2. SWARMM - AN OVERVIEW

2.1 SWARMM

The fundamental difference between SWARMM and most other simulations of air-combat is the separation of the models of the physical systems from the models of human decision making [MHA+96]. By implementing SWARMM in this way it was possible to take advantage of advances in agent oriented (AO) technologies. SWARMM is a model which has been developed to allow for an improved capability in modeling the tactical decision making aspects of air-combat. It incorporates an existing high-fidelity model of aircraft physical systems with a new agent-oriented approach to modeling human reasoning. The language chosen for the human reasoning modeling is dMARS1. A brief description of dMARS follows in section 2.2 with a more formal and detailed view available in [d'IKL+97].

2.1.1 The Tactical Reasoning Models

The tactical reasoning models allow for the simulation of the situated decision making of each of the fighter pilots as an individual dMARS agent. This agent has the capacity for situation awareness and assessment [MH95], it will exhibit team behaviour and has a notion of team responsibilities [THS97]. The pilot-agents are assigned missions which they will attempt to complete as a member of a team with assigned roles and responsibilities in that team. They will handle events as they occur in the simulated world and apply their procedural knowledge (in military parlance this can be considered as their standard operating procedures). The framework also allows for the implementation of a human performance model of the type proposed by Lloyd [Llo97] which can potentially model the effects of experience, workload, and stress.

2.1.2 The Physical Models

The physical models simulate the world and each of the aircraft flying in it with all of their associated sub-systems. This includes, but is not limited to, the aircraft performance and aerodynamics, the radars, missiles, countermeasures, passive sensors and radios. The models are implemented in C and FORTRAN and are in many cases modules developed by specialists separate from the analysts who assemble and use SWARMM.

2.2 dMARS

The distributed Multi Agent Reasoning System (dMARS) is the most mature implementation of one of the more popular architectures for agent systems; the belief-desire-intention (BDI) model. The system was a growth of the Procedural Reasoning System [GL86] developed by SRI International and represents the state of the art with respect to BDI agent systems. The dMARS system developed out of a BDI model proposed by Rao and Georgeff [RG91] based on the philosophical concepts of intentions, plans, and reasoning proposed by Bratman [Bra87]. The system allows for agents with a representation of beliefs as perceptions of its environment implemented as a relational data-base. Plans are declarative procedures which represent a set of instructions that define methods for responding to events to achieve certain ends. The plans are coded graphically (see Figure 1) which affords certain benefits to the analyst (see Section 5.3). Desires are typically descriptions of events occurring in the system as goals to be achieved. Intentions are simply the instantiated plans required to achieve some goal.


Figure 1. A Sample Mission Plan

2.3 Visualisation Tools

2.3.1 Tactical Execution Tracing

A graphical interface into the agents at run-time is provided with the dMARS system, termed the dMARS Control Interface (DCI), this allows the analyst to observe the agents' decision-making as it occurs, and is used in conjunction with a three-dimensional graphical view of the simulated world via Xcombat which is described in the next section. The analyst may also inject or retract agent goals and/or beliefs at run-time, affording a measure of external control to the situation. The DCI mechanisms and traces permit much more efficient development of studies in SWARMM than traditional scientific languages which only provide debugging and post-processing viewing.

2.3.2 Xcombat

Xcombat [Pap97] is a high-end, three dimensional graphics tool which allows the analyst to observe the physical world taking a god's-eye view or a pilot-view of the simulation as it proceeds. Xcombat can display real time information about sensors, weapons and countermeasures as well as the physical state of all of the aircraft in the simulation.

3. SWARMM - THE ADVANTAGES

3.1 Development and Control of Model Behaviour

3.1.1 Development Efficiency

SWARMM agent reasoning with dMARS operates with the agent's knowledge-base bounded by user-defined plan libraries. Each plan library contains one or more tactical plans which may be easily cloned and modified to whatever degree the analyst wishes. The rapid cloning/modification feature, apart from its revision control advantages, permits rapid development of operational behaviour.

The use of different plan libraries in SWARMM offers great flexibility by enabling scenarios with numbers of aircraft ranging from 1 vs. 1 to many vs. many, with differing aircraft roles and missions, to be easily constructed and run using the same physical model. The level of tactics included in the plan library may range from very simple behaviour to extremely complex interactions between aircraft in the same team or on the same side.

The ease and simplicity of creating scenarios and modifying behaviour translates directly to improved turn around time on tasks. The graphical nature of dMARS plans is beneficial to the analyst and to the customer, especially when displaying the nature of the tactical simulations to the latter.

3.1.2 Version Control

In operational analysis models employing a unified approach to simulating physical performance and behavioural traits (we shall term these 'traditional' models), entity behaviour typically is developed for a specific purpose. In other words tactics modelled using conventional scientific languages such as FORTRAN and C tend to be tailored for a particular study 'scenario'. As the tactics are modified to suit a new scenario, the behaviour exhibited in the original scenario may change and it may prove extremely difficult to quarantine unwanted behaviour from the new requirements for a particular scenario without undesired side-effects creeping in, even if time permits extensive testing.

Furthermore, within the course of any given study, the tactics used may change from the tactics originally agreed with the customer. For example, naive employment of a particular tactic may result in a change in the measures of effectiveness that is unrealistic. The modification of existing tactics and addition of new ones; means that version control may be problematic.

SWARMM features clear developmental benefits over traditional approaches employing conventional scientific languages such as FORTRAN and C. Apart from the ease of creating and modifying plans in the dMARS language, it is also a simple matter to reconfigure an agent's behaviour by including or deleting plans in its plan library. These plan libraries represent the agent's tactical reasoning capacity and their modularity and reconfigurability facilitate version control and segregation of model behaviour for specific requirements.

3.2 Customer Orientation

The customer-base for the types of operational analysis studies conducted by DSTO, for the Department of Defence, ranges from service and civilian public servants in force-structure evaluation, through to front-line operational units who may raise tactical analyses or underlying requests for advice.

Prior to any analyses being conducted, it is necessary to generate the desired tactical behaviour, typically this is an iterative process, in this process the customer should be consulted for feedback.

SWARMM's use of dMARS endows a distinct advantage over traditional languages when displaying and discussing tactical code with the customers. The use of dMARS and the software design, which employs military terminology in the plans and initialisation files, combine to facilitate customer orientation with the study and the important tactical modelling.

The ability to inspect the tactical reasoning structure and execution of the model, via the DCI graphical interface, endows SWARMM with distinct advantages over traditional models. Feedback from the customer is also achieved efficiently via the DCI, as rapid changes to the reasoning can be incorporated as required.

4. SWARMM - HOW IT IS USED

4.1 Requirements Analysis

Dependant upon the customer's origins and the nature and timescale of the study, there will be differing emphasis on the tactical and physical aspects of the simulation. This paper concentrates on the tactical requirements of the study since it is in this area that SWARMM has a major advantage over traditional models.

SWARMM analysts work closely with the client at the outset to determine the correct mix of tactical and physical performance required, although the analyst's domain experience is used to determine the optimum means of addressing the study scope with SWARMM. In the case of simple studies, such as tactical defence options against missiles, the analyst will progress from a simple brief.

Once tactical simulation requirements have been scoped, the next step is to determine if existing dMARS tactical plans are adequate, modifications to existing plans will suffice or if new plans may be required. It is in this area that relatively rapid development is possible and where SWARMM excels.

A relational database has been developed which permits clear appraisal of the existing tactical libraries, including the querying of the database by plan keywords, or by tactical phrases. Use of the database clearly facilitates the planning stage of the study, including rapid determination of the tactical plan infrastructure required.

4.2 Study Planning

The planning stage of the study is closely intertwined with the requirements analysis. The analyst must be familiar with the nature of SWARMM's tactical reasoning model and the physical system modules. An example would be a study assessing missile performance at the edges of its launch envelope, recognising the need to override some pseudo-tactical control logic in the physical model which might otherwise prevent a missile firing. This pseudo-tactical logic remains in the physical model for reasons of development priority and would usually be transparent to the user, however a specific study such as the example given would require modification, in which case a specific version of the SWARMM model is created that is tailored to the requirements analysis. There is a one to one correspondence between the study scenario and the SWARMM executable. The is carried out by analysing the variations in important measures of effectiveness produced by variations in either tactical doctrine applied or changes in physical systems or both. These variations are considered as unique "cases" of the scenario and are determined by variations in the data input to SWARMM.

A "study" that requires more than one version of the executable code (eg. two versions of a physical model) may be considered as two separate but related studies.

4.3 Model Creation

The flexibility inherent in SWARMM provides considerable advantages, however, the construction of the individual components of SWARMM needs to be managed carefully. A central repository is established to contain the version of SWARMM created for the study. Each developer/analyst can create a local version of SWARMM based on a combination of the local source code and the source code contained within the central repository. When the code developed has been thoroughly tested and approved, the appropriate source is moved to the central repository for inclusion in the study version of SWARMM.

The plans required for SWARMM fall into two categories. Firstly, there are those associated with the framework of SWARMM - the passing of messages between the reasoning model and the physical model, the deliberation about new information and so on - which are required for the proper functioning of SWARMM and are study independent. Secondly, there are the tactical plans detailing what actions to perform in a particular situation, these are highly study dependent.

The plan database is accessed to determine what plans from previous studies are appropriate for the study under consideration. The "framework" plans are usually taken as a whole from a previous study as they are study independent. The tactical plans from previous studies are examined to assist in designing the plans for this study but in general are not recycled.

The physical models used within SWARMM are usually study independent and can be carried over from a previous study. For example, the model of a missile is the same regardless of the study in question (unless, of course, the study examines the changes in operational effectiveness due to modification of the missile).

The variations that are required for a study involve changes to the data input into a physical model, for example, the number or type of missiles carried. For each study "case" a directory is created with the appropriate input data and a link to the SWARMM executable. The measures of effectiveness (MoE) also introduce specific post-processing requirements on SWARMM and may require appropriate changes to model output data.

4.4 Study Variations and Perturbations

The outcome of an air combat engagement may be strongly influenced by the relative initial geometry of the combatants and careful consideration must be given to ensure that these initial positions of the aircraft do not unduly influence the simulation results. Hence, for some studies the starting positions of the aircraft are randomly perturbed.

Other variations may include the randomised initial directions of radar search beams, of infra-red flare velocity disturbances, and of countermeasure effectiveness evaluations. All of these pseudo-random variations are inherent to the physical model and required for Monte Carlo statistical analysis, however the analyst may wish to concentrate upon certain aspects for a given study, by modifying the basic behaviour.

Whilst a number of physical models used within SWARMM incorporate random effects, the radar model in particular uses a random number generated to model the randomness associated with radar detection. The seed value used to start the random number generator can be explicitly specified or selected based on the time of execution.

SWARMM is a deterministic model where the model output for a specific case can be regenerated by specifying the seed for the random number generator and the input data for that case. Any perturbations in platform initial geometry are controlled by the random seed.

Study results are obtained by executing the SWARMM model a large number of times.

5. SWARMM - AN EXAMPLE

5.1 Study Scenario

This study aims to evaluate the effectiveness of three types of missile evasion manoeuvres (Manoeuvre A, Manoeuvre B and Manoeuvre C) that a pilot might consider when constrained for various. While this example is a relatively simple 1 vs 1 scenario, it is illustrative of the way in which SWARMM is used.

Each of the manoeuvres to be evaluated requires a new plan or set of plans to be written which will replace or override the standard missile evasion manoeuvres in the target aircraft's plan library. In addition, plans may need to be written or modified in order to ensure that the target reacts to the missile launch at the appropriate time and begins the evasion manoeuvre in the appropriate way. In order to evaluate each manoeuvre separately, each manoeuvre type becomes a separate scenario.

As stated previously, the physical model contains pseudo-random variations that influence radar detection of the target as well as the targets observation of a possible missile firing, whilst the starting positions of the aircraft are also randomly perturbed to counter the possible influence that the initial positions of the aircraft may have upon the simulation results. A large number of runs may then be carried out and Monte Carlo statistical analysis performed (see section 5.3.1).

5.2 Study Cases


Table 1. Study cases to be considered.

A range of starting conditions for each scenario are considered, Table 1: these are used to examine the sensitivity of the manoeuvre effectiveness to changing the parameters defining one or more of the manoeuvres. This would result in a further set of cases to be investigated and an expansion in the table.

As an example, for any target/launcher geometry there is a corresponding maximum range, Rmax, at which the missile is able to intercept a non manoeuvring target. As the launch ranges reduces from Rmax the missile is increasingly able to intercept a target that is manoeuvring. Therefore, the percentage of Rmax at which the missile is launched influences the effectiveness of any evasion manoeuvre carried out by the target. A sensitivity analysis of the launch range upon manoeuvre effectiveness may be carried out by varying the percentage of Rmax at which the launch aircraft is allowed to fire its missile. This further expands the number of cases to be considered and, as more parameters are added to those to be analysed, the number of cases to be investigated rapidly becomes excessive.

It may therefore be necessary to perform a mini-study to evaluate the sensitivity of the manoeuvre effectiveness due to a change in selected parameter values. A decision as to what parameters are likely to have greatest influence upon the results may then be made to limit the set of cases to be investigated. A more comprehensive study of the reduced set of cases may then be carried out.

5.3 Analysis of behaviour

5.3.1 Post-run processing

A number of post-run processing tools have been developed for analysis of both individual runs and for Monte Carlo set of runs.

Post-run processing of output from Monte Carlo runs may be carried out using the tool studProcess [Smi97]. Using studProcess it is possible to correlate various parameters output to file by SWARMM and create histograms of those parameters. For example, there are a variety of conditions that will prevent a missile from successfully intercepting a target and these are output to file [Tur97]. It is therefore a simple matter to post process results obtained from a large number of Monte Carlo runs and collate the missile failure reasons against parameters such as range at launch, target delay in reacting to missile launch, etc (see Figure 2). This will enable the analyst to determine the effect of such parameters upon the success of the evasion manoeuvre.


Figure 2. Histogram of Missile End Conditions with Range at Launch for Case 1

However, to explain why any manoeuvre was or was not successful it may be necessary to examine one or more representative runs in more detail. For this example, missile performance parameter time histories may be post processed for any aircraft in an individual run. These time histories may then be used in conjunction with run-time visualisation tools to gain an insight into the effect the manoeuvre has upon the missile guidance system.

5.3.2 Run-time analysis

As discussed in Section 2.3.1, the DCI gives the analyst the facility to select any agent and to view some or all of the plans as they are invoked and to proceed through the reasoning process step by step. Thus, the DCI, in conjunction with the Xcombat visualisation tool, enables the analyst to equate the decision-making process with the observed behaviour. The analyst is then better able to identify the tactical decisions which lead to an agent's response and to detect any deficiencies in the tactics. The relevant plans are then easily modified.

6. STUDY REPORTING

The style and content of reporting must be tailored to the target audience which, in the case of operational units which may have tasked a SWARMM tactical analysis, there is a need to generate results which reflect the needs of these customers to have data and supporting descriptions phrased in language familiar to them. The level of detail employed in the report must be commensurate with demonstrating the analysts' knowledge of the domain, both the performance of physical systems and the tactical environment which governs the employment of such systems.

An example of operationally focussed client reporting is shown Figure 3 where tactics are simply presented in terms of relative effectiveness and tactical issues affecting the results would be described in appropriate terms.


Figure 3. Effectiveness of missile evasion manoeuvres

For studies which have been tasked by other areas of Defence, the reporting style will be closer to the nature of a scientific report. There will be more justification of study cases, the procedures adopted on starting conditions, the scope of the study and necessary assumptions. All of the above are, of course, relevant to the operational unit, however it may be sufficient to use some pointers to these, in the knowledge that a mutual understanding must exist at the working level.

7. RELATED DEVELOPMENTS

SWARMM has spawned a number of other technology developments with similar characteristics and is being developed further. The tactical libraries are being expanded to meet the requirements of the air force and a variant for use in a manned simulator as a provider of computer generated forces is being developed [MHA+97]. As simulations become larger and more complex there is a trend toward distribution of the software to aid in the management of the system. AOD is presently developing a distributed simulation environment to ease some of the software engineering problems which are often encountered. The distributed version of SWARMM (known as dSWARMM) is in the analysis and design phase.

8. SUMMARY AND CONCLUSIONS

SWARMM is a dramatic departure from existing simulation codes. It uses sophisticated BDI agents to model human decision making and a novel approach to separating models of tactical decision making. The policy of selecting a software solution best suited to the particular problem has ensured that SWARMM is efficient and flexible.

SWARMM is now a mature system in everyday use within DSTO for the study of military OR issues. By fostering close interactions with operational squadrons and by selecting technologies which allow for rapid response and ease of transfer of information AOD are able to answer complex analytical problems in a timely and relevant manner. In the rapidly changing and increasingly more complex world of modern military systems there is a need for modelling and simulation to stay at the leading edge of technology growth. SWARMM is a leading edge example of modern OR techniques which can be employed to address effectiveness studies in the highly complex military environment.

SWARMM emphasises the importance of the human operator as a part of the total system and allows for the incorporation of human factors issues such as workload, stress, and skill-level into the evaluation of tactics and hardware.

Experience to date has demonstrated the advantages of employing BDI agent technology for tactical behaviour modelling, with traditional mathematical models of physical systems and the environment. The combined dMARS/physical system approach has proven advantages over traditional equivalent models during model design and coding, development testing, separation of study variations, and study execution and post-processing. These advantages magnify when the complexity and scope of tactical modelling increase with the scale of the required study.

Future developments of SWARMM will be extended to include greater representation of human performance in the agents, and applied at the campaign as well as the engagement and mission levels.

9. ACKNOWLEDGMENTS

The authors are grateful to Mr Graeme Murray and Mr David Glenny for their reading and comments on this paper.

REFERENCES:

[Bra87]  M. Bratman. Intentions, Plans, and Practical Reason. Harvard University Press, Cambridge MA, 1987
[d'IKL+97] M. d'Inverno, D. Kinny, M. Luck, and M. Wooldridge. A Formal Specification of dMARS. In Pre-Proceedings of the Fourth International Workshop on Theories, Architectures and Languages, 1997.
[GL86] M. Georgeff and A. Lansky. Procedural Knowledge. In Proceedings of the IEEE Special Issue on Knowledge Representation, volume 74, pages 1383-1398, 1986.
[Llo97] I. Lloyd. Simulating Human Characteristics for Operational Studies. DSTO Research Report DSTO-RR-0098, 1997.
[MHA+96] D. McIlroy and C. Heinze, Air Combat Tactics Implementation in the Smart Whole Air Mission Model (SWARMM). In Proceedings of the First International SimTecT Conference, Melbourne Australia, March 1996.
[MHA+97] D. McIlroy, C. Heinze, D. Appla, P. Busetta, G. Tidhar and A. Rao. Towards Credible Computer-Generated Forces. In Proceedings of The Simulation Technology and Training Conference (SimTecT), Canberra, Australia, March 1997.
[Pap97] M. Papasimeon. The Xcombat Users Guide. Air Operations Division working paper, unclassified, 1997.
[RG91] A. Rao and M. Georgeff. Modeling rational agents with a BDI architecture. In J. Allen, R. Fikes, and E. Sandewall, editors, Proceedings of the Second International Conference on Principles of Knowledge Representation and Reasoning., pages 473-484. Morgan Kaufmann Publishers, San Mateo, CA, 1991.
[Smi97]B. Smith. Post-Processing Software for PACAUS and SWARMM. Air Operations Division working paper, unclassified, 1997.
[THS97] G. Tidhar, C. Heinze, and M. Selvestrel. Flying Together: Modeling Air-Mission Teams. In preparation. To be published in Apllied Intelligence , Kluwer Academic Publishers, Boston, 1997.
[Tur97] M. Turner, PACAUS/SWARMM: Missile Model Success/Failure Flag Meanings. Air Operations Division working paper, unclassified, 1997.

The Distributed Multi-Agent Reasoning System (dMARS) is a product of the Australian Artificial Intelligence Institute [ http://www.aaii.oz.au/proj/dMARS-prod-brief.html].

- -