Introduction

The Australian Defence Force is undertaking a study of the current and future capabilities of the Land force, the potential benefits and limitations these capabilities bring to the Land force, and the effects produced by various combinations of these capabilities at the tactical and operational levels (Commonwealth of Australia, 2004). To guide the development of Land force capabilities as its underlying requirements, needs and imperatives evolve over a 30-year timeframe in an ever-changing geopolitical environment, the ADF 'must be able to articulate concepts and effects to meet the range of potential conflicts, operational environments and Military Strategic Objectives that are currently uncertain' (Curtis and Dortmans, 2001). This paper studies the relationship between the capabilities of the Land force and the fundamental characteristics, attributes and skills inherent to parties within the force, which enable, facilitate and support the production, application and sustainment of effects in the tactical Land environment.

Perhaps the first simulated study of Land's capabilities and the relationship of these capabilities to a conceptual model of the Army as a set of enabling core skills (Curtis and Dortmans, 2001) was conducted by Shine (2005). In this study, a conceptual model of the Army was developed using an agent-based approach. The notional Army was described in terms of a number of interacting entities, weapon systems or assets, each endowed with physical attributes, behaviours and simple reasoning. These entities were then implemented as agents in the combat simulator MANA (Lauren et al, 2001). To link the agent-based model back to the notional Army, each different type of agent implemented in MANA was related back to one or more Army core skill. For example, the agent called better-firepower was related to the Army core skill engagement. Forces of the same types of agents and limited mixes of two different agents were then matched against each other in battle. This approach was limited to defining specific instances of relationships between core skills and a discrete number of different types of agents in MANA, rather than general relationships based on a sliding scale of agent capabilities. Furthermore, because the core skills and the behaviours in MANA were related to each other by subjective interpretation rather than as characteristics inherent to agents, the appropriateness of the relationships is questionable.

Instead of using Shine's (2005) approach, we propose a set of six core skills: communication, detection, lethality, mobility, protection and sustainment. We rate agents' abilities in the core skills on an ordinal scale between 1 and 10. This implementation of core skills leads to an intractably huge space of possible combined arms teams. To search this space, we follow an approach used by Baker et al. (2004) in a related preliminary study and apply an evolutionary search algorithm (Spears et al, 1993). Games theory (Von Neumann and Morgenstern, 1944; Nash, 1950, 1951) is applied to provide a quantitative valuation of the consequences of adding, removing and replacing capabilities within combined arms teams. This approach measures the value of potential combined arms teams loosely based on the sensitivity of the model to changes within sets of candidate options. A discussion of this study's findings and its conclusions are presented in the final two sections.

Agent model

Agents are characterized by a set of six attributes that represent the agents' capabilities in communication, detection, lethality, mobility, protection and sustainment. These attributes, rated on the ordinal scale with a minimum value of one and maximum value of 10, affect how agents interact and the behaviours they exhibit as described below. These behaviours are designed to be as simple enough to model and implement, yet sufficiently complex to give reasonable results for analysis. Simulations are conducted in a discrete-time turn-based environment so that each agent is given the opportunity to make use of its attributes at least once per simulated turn and potentially more (depending on the sustainment attribute). We refer to the agent's use of an attribute as an action or an agent's action.

• Communication. Agents attempt to communicate enemy sightings to their force. An agent's communication attribute defines the maximum number of distinct communications that can be processed in one action. This value, when divided by 10, also defines the probability that each independent communication is successfully transmitted. The agent is aware of the success or failure of each attempted communication and will retransmit the same report if it is able to. That is, if the number of communication attempts is still below the maximum number defined by its communications attribute. Upload of information is assumed to be either a manual or automatic process over a carrier susceptible to failure; for example, sending a Variable Messaging Format (VMF) transmission over a combat net radio or Battle Management System (BMS) by line-of-sight or range extension via relay. Download of information is assumed to be an automatic process over a carrier that is not susceptible to failure in open terrain; for example, receiving blue-force-tracker information and known red-force locations facilitated by a BMS over a secure satellite.
• Detection. An agent's detection attribute defines the maximum range over which enemy agents can be sensed. This value, when divided by 10, also defines the probability that enemy agents are successfully detected between one-thirds maximum and maximum ranges in a single action. At ranges between zero and one-thirds of the maximum range, enemy agents are always successfully detected. For simplicity, no distinction is made between detecting, identifying and recognizing. Agents are either detected or not. Agents maintain both local and force-level situational awareness, and thus may be aware of enemy targets beyond their own detection range. Hence, agents may act as spotters or forward observers for their own force.
• Lethality. An agent's lethality attribute defines the maximum range over which enemy agents can be prosecuted. This value, when divided by 10, also defines the probability that enemy agents are successfully engaged between two-thirds maximum and maximum ranges, that is, the probability that the enemy agent is damaged or destroyed in a single action. At ranges between zero and two-thirds of the maximum range, enemy agents are always successfully engaged.
• Mobility. An agent's mobility attribute defines the maximum range over which the agent can manoeuvre in a single action. Terrain is not modelled in this study. How agents move and the movement algorithm itself is omitted for brevity, but can be found in Wheeler (2005). Briefly, movement is designed such that the blue and red forces advance as a coherent group with the objective to capture and hold the middle of the battlefield, where each force is initialized on opposing sides.
• Protection. An agent's protection attribute defines the number of times the agent can sustain damage before being destroyed. For simplicity, agents' capabilities are not diminished in any way by damage. Agents are either destroyed or functional. The protection attribute can be interpreted as the amount of physical armour carried by the agent, as a numerical size representing an agent comprised of more than a single individual or a combination of the two.
• Sustainment. An agent's sustainment defines the total number of times that it may scan for enemies and operate its radio or employ its weapon. This effectively gives the agent additional actions in detection and communication or lethality or both, in a single simulated turn. The agent randomly chooses between the two options, but will not employ its weapons if no enemy targets are detected. Sustainment is best interpreted as enabling agents to operate equipment at maximum efficiency. As such, it represents a rate of use which is likely to denote a combination of training and experience, characteristics of the hardware and software systems, as well as the level of combat service support and resupply.

For brevity, details of the combat simulator beyond those supplied above are omitted, but can be found in Wheeler (2005).

Empirical results

Parameter space exploration

In this section, a limited study of the interactions between the six attributes is conducted. R=500 simulations are run between two distinct forces in which each force is enhanced, from an initial balanced force in which each agent has P/6=3 points for each attribute, by adding a single point to two of the attributes of every agent in the force. There are 21 ways to distribute two points among six attributes, including variations in which the two points are added to the same attribute. R simulations are conducted for each of the 210 unique pairs of the 21 possible forces to calculate the fitness of the each simulation. To calculate the fitness of a simulation, B=20 battles are conducted. Hence, BR=10 000 independent battles in total are conducted for each of the 210 unique pairs.

Simulation results are tabulated in Figure 1. In this figure, the mean fitness is entered into the matrix to one significant digit. Rounding the results to one digit is entirely safe when it is observed that the greatest variance over all results is approximately 0.50 so that the standard deviation of all results is strictly less than 0.71. The t-statistic for a two-sided 99.9% confidence interval about the sample mean at R=500 degrees of freedom is approximately 3.3101, and in the results the largest interval for the sample means is 0.104706. Hence, we are certain within reasonable doubt that all results are statistically valid. When rounding results to 0, the signs + and - are preserved for interest. For example, the fitness of lethality against protection+communication is approximately -0.445750. This value is recorded as -0 in Figure 1, denoting that a force enhanced by augmenting agent's lethality attributes twice defeats a force enhanced by augmenting agent's protection and communication attributes once each, and the defeat occurs with a fitness value strictly more than -0.5 and strictly less than 0.

Figure 1.

Simulation results.

Full figure and legend (91K)

Mobility is not displayed in Figure 1. Without exception, forces with enhanced mobility performed poorly. This is a direct consequence of the scenario and not necessarily a comment on the usefulness of enhanced mobility in general. The scenario under study specifically and intentionally models a battle in which two opposing forces encounter each respective opposing force without prior warning. The scenario is developed such that each side has sufficient time to form as a combined arms team, hence reducing the potential for ambush or strike in the enemy territory. Enhanced mobility in this study merely means that the force attains the goal point at the centre of the battlefield before its opponent. As a consequence, the point at which combat occurs is moved slightly from the centre of the battlefield into the enemy territory. For all practical intents and purposes, this scenario is insufficient to determine the relative benefits of enhanced mobility. However, such a study could be conducted by assigning a tactical benefit and liability, perhaps in combat and detection range, at different regions of the battlefield. For example, a hill could be marked as a tactical advantage and a valley as a tactical disadvantage. It is then important for each force to defend regions with a high tactical advantage and attack regions with a low tactical advantage. Mobility is critical for such manoeuvres.

Simulation results are analysed using the analytic hierarchy process (AHP) (Saaty, 1977, 1980), a method of multiple pairwise comparisons for a finite number n of options. We translate the fitness scores f(i,j)=-f(j,i), which describe the fitness of the force i against the force j, for the AHP using

where the m_i,j are the entries of the matrix M of multiple pairwise comparisons to be analysed. The normalized principle right eigenvector w of M approximates the priorities of the options (see Figure 2). Mobility is included in this figure for completeness.

Figure 2.

Comparative importance.

Full figure and legend (47K)

In Figures 1 and 2, the attributes broadly fall into three levels of effectiveness. Lethality, protection and sustainment are most important to a force. Next, communication and detection is important. Mobility is the least effective attribute in this scenario. However, the ways in which attributes combine are interesting. For example, lethality alone is only ranked seventh in importance, but the combinations of lethality with protection and sustainment are ranked first and second, respectively. The reasons for these results are as follows.

• Lethality. Ultimately, a force must be able to inflict damage on the enemy to defeat it. Lethality directly promotes a force's ability to prosecute the enemy and is useful even when the force has poor detection and communication. Any detection and communication that does take place is acted upon with reasonable confidence of inflicting damage upon the enemy.
• Protection. The ability to survive contact with the enemy is paramount in all combat. High scores in all other attributes fail to negate the simple necessity to remain operational under fire.
• Sustainment. Sustainment increases an agent's ability to detect, communicate and prosecute. However, it is not a panacea and acts only to supplement sufficiently adequate natural attributes in agents and to minimize deficiencies in one or two of the agent's attributes. The outcomes of simulations are sensitive to the relationship between sustainment and the agent's other attributes. For example, a base score of 2 substantially reduces the efficiency of sustainment, while a base score of 4 substantially increases the efficiency of sustainment.
• Communication. Informing one's allies of the location of enemy agents is important to defeat those agents as quickly as possible. Communications is a second-order influence on the outcome of combat and is of less importance to the force if agents possess sufficient firepower and protection to defeat the enemy without assistance.
• Detection. A force must be able to detect enemy agents in order to prosecute them. However, being able to detect the enemy does not mean that the agent has the firepower to engage and destroy that enemy, the communications to alert the force to the enemy's presence or the protection to survive the encounter. For these reasons, detection is of secondary importance.
• Mobility. Mobility does little to assist a force in this scenario.

The AHP defines a consistency ratio to measure the accuracy of recorded results, in the matrix M, against random data. In this study, a consistency ratio of approximately 0.008 or 0.8% is obtained. This is substantially lower than the recommended 0.10 or 10% inconsistency limit and describes an overwhelming level of consistency in the results.

The consistency ratio of M is at odds with Shine's (2005) study in which a broad skill set defeats engagement and decision making, engagement and decision making defeat information collection and movement, and information collection and movement in turn defeat a broad skill set. Our consistency result suggests that there exists a single best force, composed of identical agents, which defeats all other possible forces which are likewise composed of identical agents. However, it is easy to observe from Figure 1 that no such force exists within the proposed 15 forces. Lethality coupled with protection is the preferred force option, but a two-person zero-sum normal-game G with payoff matrix as presented in Figure 1, and with additional payoff values of 0 on the leading diagonal (top left to bottom right), has no saddle point solution which would describe the single best force solution to the game. No inference can be made about forces other than those options examined in this section.

Evolved strategies

In the previous section, we present an overly simplistic model designed only to study simple homogeneous forces (comprised of agents with identical capabilities), which do not represent combined arms teams. To examine such teams, we relax the constraint on the blue force and allow it to evolve against various red forces using a genetic algorithm. This allows us to examine true combined arms teams or heterogenous forces (comprised of agents with different capabilities), in which agents assume diverse roles and functions within their teams.

In this study, the foundation of combined arms teams are single units or agents. Having taken an agent as a basic building block, we define the concept of a combined arms team as a collection of N agents and a population as a collection of T teams. A population then describes a collection of candidate solutions for combined arms teams.

Let there be P points to distribute between K attributes, where PK. Then the total number of distinct agents, in the set of possible agents having no attribute less than one, is implicitly given as the solution to

where x_i, i=1...K, are positive integers.

Counting the number v of solutions to the equation above is equivalent to counting the number of possible ways to allocate P identical objects to K distinct bins. This problem has solution

The number of distinct combined arms teams composed of N agents, or the number of ordered collections of N agents in which agents may appear more than once, is then

The total number of distinct combined arms teams is so intractably huge for the parameter sets we use that it could not be computed using the equation above on a Pentium IV processor. Hence, to search this space for desirable combined arms teams, matching some set number of criteria yet to be defined, a heuristic algorithm is necessary. The heuristic algorithm we use is the genetic algorithm.

We base our genetic algorithm loosely on Goldberg (1989) and Holland's (1975) seminal work on genetic algorithms and operators, but observe that our particular adaptation is a new and unique interpretation of their original formulations. For brevity, details of our implementation are omitted. It suffices to say here that the algorithm acts only on agents and not binary or real coded strings, and that the algorithm contains an elitist operator, which stores the fittest solution found, in addition to the three standard operators: selection, crossover and mutation. In this study, fitness is calculated as the average difference in the number of casualties sustained by each force over B independent battles. A complete explanation of our algorithm can be found in Wheeler (2005).

Two cases are considered:

1. The blue force evolves against a balanced red force.
2. The blue force evolves against a randomly generated red force.

R=500 independent replications of each of the two cases above are conducted. The mean fitness of the blue combined arms teams across the population or gene pool of T=10 combined arms teams in the genetic algorithm and the fitness of the best individual combined arms team in the population are recorded for each replication. The mean and standard deviation across the R replications are displayed in Figures 3 and 4.

Figure 3.

Evolution against balanced force.

Full figure and legend (36K)

Figure 4.

Evolution against random force.

Full figure and legend (40K)

Interestingly, the average mean fitness for the blue force at generation g=0 is approximately -0.335 for case (1). This indicates that a randomly generated force, as is the blue force at g=0, is defeated by a balanced force.

Let A_y, where y{Lethality (L), Mobility (M), Protection (P), Detection (D), Communication (C) and Sustainment (S)}, denote the random variable which represents the outcome of an agent's y attribute value at the completion of a simulation. Then the Shannon entropy for the agent's attributes is

The mean attribute values across all A agents in the T combined arms teams at the completion of each of the R simulations is presented in Table 1, together with the standard deviation for the mean and the base b=e Shannon entropy, as calculated by numerical approximation.

Table 1 - Mean attribute values.

Full table

The mean value for the attributes L, M, P, D, C and S in Table 1 can be interpreted only as the average total investment, in points per attribute, of a combined arms team per agent. The mean in no way denotes a representative agent for the combined arms team. The Shannon entropy measure is a better measure of the volatility in agents' attributes than the standard deviation because the standard deviation for the mean becomes arbitrarily small when sufficiently large numbers of simulations are conducted. For the sake of comparison, the uniform distribution over the Integers between one and 10 has a Shannon Entropy of approximately 2.303 nats and any unit point distribution has a Shannon Entropy of 0 nats exactly.

The results of Table 1 are surprising in light of the analysis of the previous section. In particular, there is an unexpectedly high average total investment in detection at the cost of protection. One surmises that, in combined arms teams where agents evolve to counteract any deficiency held by the team as a whole and diversify to develop specialities, detection is the most universally important attribute, while protection is only more important than mobility. However, the high variability in the mean for protection in relation to the other attributes indicates a greater range of values for that attribute. Agents rely on high values in other attributes and diverse combinations of agents with special functions within the team to offset lower survivability.

Finally, a blue force is evolved against a semi-balanced red force. This red force consists of agents initialized as balanced with attribute values of P/6=3 and then augmented twice, in a similar manner as the agents in the previous section, that is, by adding 1 unit to two attributes chosen at random. However, in this semi-balanced force, the agents are independently generated. Hence, the red force is not composed of identical agents as are the forces in the previous section. The average fitness of the genetic algorithm upon termination of the R simulations is approximately 1.12, with a standard deviation of 0.53, for the mean across all T combined arms teams and 2.50, with a standard deviation of 0.62, for the best combined arms team. This result is interesting in that agents in the blue force have P=18 attribute points in total and agents in the semi-balanced red force have 21 attribute points, approximately a 17% increase in capability over the blue force. This illustrates the very real value in diversity in the blue force and the importance of the right composition of agents.

Coevolved strategies

Up to this point, we have discussed the application of a genetic algorithm to evolve a single population, for the blue force say, against a static opposing-force team, the red force say. Such an implementation is useful in determining a blue force combined arms team best suited to defeat a given red force team. However, one may equally well ask what happens when the red force is also simultaneously evolved against the blue force. A coevolving implementation of the genetic algorithm involving two populations is required. This two-population genetic algorithm is a simple extension on the single-population genetic algorithm that interlaces two single-population genetic algorithms to evolve against each other.

In Figure 5, the average fitness of the blue and red populations is displayed for a single simulation of the two-population genetic algorithm. The average fitness is calculated using R=500 independent calculations for the fitness, where each calculation simulates B=20 independent battles, of each combined arms team in each population against their opposing force's elected best combined arms team. The standard deviations at all points cannot be displayed using error bars because the largest deviation is approximately 0.003, which is too small to depict.

Figure 5.

Coevolved average fitness.

Full figure and legend (27K)

In Figure 5, it is interesting to note the interplay between the red and blue populations. As a trend, the two data sets seem to oscillate about the x-axis and each force opposes or counter-balances the efforts of the other. For example, near iteration 30 the red force has exploited a weakness in the blue force, but near iteration 36 the blue force adapts to the red force and then dominates the battlefield until iteration 47.

To examine the actual composition of the combined arms teams resulting from coevolution, a two-person normal-form game G is constructed. The red and blue populations at the end of three independent coevolution processes are recorded. The six strategies for these combined arms teams are available to each player irrespective of whether or not they originated as a red or blue strategy. If a strategy for a combined arms team is a good strategy for the blue force, then it is an equally good strategy for the red team and vice versa. No constraints are imposed on either player as to which of the six strategies to employ. The combined arms teams generated will differ each time this analysis is conducted. The actual teams used in this example can be found in Wheeler (2005).

To populate the normal-form matrix M, we define the payoff function P(i,j), where i and j belong to a finite set of strategies, as the average fitness, attained over R=500 independent fitness calculations, where each fitness calculation represents the average outcome of B=20 battle simulations. These values are presented to two decimal points as the zero-sum normal-form matrix M below, where the leading diagonal is set to 0 and m_i,j=-m_j,i is enforced. The 99.9% confidence intervals for the entries m_i,j using a t-statistic with R-1 degrees of freedom are strictly less that 0.1. Hence, the confidence intervals for the means are quite small.

The reasons why certain combined arms teams defeat others are not easily identified. However, analysing the game G helps us to understand the intrinsic value of combined arms teams.

Strategies two and four in M are strictly dominated, and in the reduced matrix there is no 'best' combined arms team because the matrix contains no saddle point solution. The combined arms teams interact with each other to defy transitivity. Team 1 defeats team 3 and team 3 defeats team 5. However, team 5 also defeats team 1. This behaviour has immediate implications for combined arms teams. In particular, future concepts for combined arms teaming should emphasize the development of agile and rapidly re-configurable forces. This section implicitly assumes that both sides of a conflict can co-adapt to each other with equal ease. This need not necessarily be the case. In reality, the rate at which a force can adapt to the opposition, and not just the strategies they have built up for combined arms teaming, has a significant impact on the outcome of conflict. Hence, the value in capability comes not just from the number of different options one has at hand but also from the rate at which different options can be utilized.

The zero-sum normal form game G with payoff matrix M has the solution S of all linear convex combinations of

This means that any two percentages, totalling to 100%, of the first and second rows of S added together will also be a valid solution.

The set of six combined arms teams analysed are used almost arbitrarily for the sake of example. We do not envisage candidate combined arms teams to be generated according to a genetic algorithm in non-theoretical studies. In reality, various options for combined arms teams might be generated to meet a number of physical constraints based on resources, existing capabilities or future acquisitions. However, once a set of candidate combined arms teams is constructed, games theory provides a number of useful insights and techniques of analysis.

For any zero-sum normal form, one can explore the consequences for either player of losing one or more strategies from their respective sets of strategies. Loss of a strategy has the potential to influence the value of the game. Each player can at best maintain the value of the game and is placed at a disadvantage whenever the strategies lost are a part of a mixed strategy solution of the game. In our example, not all of the combined arms teams are equally important in guaranteeing the value of the game. Teams 2 and 4 are unnecessary because they are not used in any solution of the game. Only one of teams 5 and 6 are needed to guarantee the value of the game so that if one of teams 5 and 6 are removed from either player's strategies then that player can maintain the value of the game. On the other hand, teams 1 and 3 are essential as both of these teams appear in every mixed strategy solution of the game. When the combined arms team 1 is removed from player 1's set of strategies the value of the game decreases by 0.0472, and when the combined arms team 3 is removed from player 1's set of strategies the value of the game decreases by 0.0887.

The value of the game is a useful way of determining the relative benefits and costs of various sets of combined arms teams. Options for combined arms teams can be added and removed and the impact on the value of the game against a baseline can be measured. However, there are some limitations to this approach. For example, suppose that the payoffs for some strategy approach that of strategy from below, such that strategy is dominated by strategy by an arbitrarily small amount >0. Then, removal of strategy from player 1's options has little impact on the value of the game because strategy can be substituted for strategy and decreases the value of the game by at most . Suppose that strategy is originally removed from player 1's options because the platforms or assets that comprise this strategy are destroyed. Then it may also be that strategy , or any other strategy, contains many of the same platforms as strategy and thus cannot in all practicality act as a substitute for strategy . The importance of strategies to each player therein lies both in the payoffs gained by employing those strategies and also in the abstract notion of a strategy denoting a genuine reserve or back-up capability that can be fielded if required and if assets are still available.

In many cases, it is reasonable to assess the impact on the value of the game of substituting not entire combined arms teams, but individual assets of combined arms teams. Suppose that four distinct combined arms teams appear in the optimal strategy of the game and that an agent representing a 120 mm mortar appears in only one of these strategies. It may be costly to maintain a 120 mm mortar capability simply for one potential combined arms team. Thus, we may wish to replace that capability by, say, a 105 mm artillery platform and then judge whether or not the decrease (or increase) in the value of the game is acceptable when weighed against the perceived cost of fielding a 120 mm mortar system. Likewise, the relative merits of integrating a number of different weapons systems into a given combined arms team can be measured.

Discussion

In this study, it is difficult to interpret the agents' attributes in terms of physical systems. To address this problem, agents' attributes could be implemented to represent real-life capabilities and the model scaled appropriately. For example, an attribute of 1 for lethality could be set to denote small-arms, 2 to denote a 81 mm artillery using standard HE point detonating rounds, 3 to denote a ground-based air-defence capability similar to a shoulder-launched ground-to-air missile, and so on. With this approach, restrictions on the use of weapons, realistic weapons effects, realistic intelligence, surveillance and reconnaissance, and range-based and line-of-sight restricted communications can be incorporated into the model. Of course, this also requires that terrain, vegetation and elevation be added to the battlefield.

Currently, the behaviours of agents in battles are fixed. Agents, irrespective of their attributes, employ the same movement, engagement, detection, communication and sustainment rules. Potentially, adaptive learning can be introduced into the model to enable agents to choose from a set of candidate behaviours and then learn from those choices. Agents then adopt various specialized roles or functions according to their attributes to accomplish simple tasks, such as taking and holding the centre of the battlefield. This approach has the potential to implement and use sophisticated tactics, and greatly enhances the usefulness and realism of the model. It is also possible that games theory solutions be used within the combat model, genetic algorithm and learning algorithms. Such a hybrid approach facilitates a more practical valuation of the combined arms teams and the agents within those teams than does the simple fitness value based on casualties.

Future applications of this research include a study of combined arms teams' sizes and the investment of resources across the teams. In such a study, force commanders select the number of agents to deploy and allocate attribute values across those agents up to a fixed amount. Such a study provides insights into the benefits and limitations of deploying large combined arms teams against smaller teams and the trade-off between numerical superiority against enhanced capability.

Finally, extensions to this study include introducing a number of different forces into the battlefield, each with tendencies, behaviours and objectives as well as Command and Control structures modelled using embedded social networks. A cooperative game emerges as the different factions ally to accomplish set goals.

Conclusions

This study has presented a foundation for the analysis of the costs, benefits and limitations inherent to combined arms teaming in a simulated environment. In our approach, six key attributes of combined arms teams, lethality, mobility, protection, detection, communication and sustainment, were integrated into an agent-based combat model. Interactions between agents were then studied by assigning attributes to the entire force and also to agents independently. At the force level, it was broadly concluded that lethality, sustainment and protection had the greatest influence on the outcome of combat, with detection and communication having a lesser impact. At the agent level, it was found that the total investment in detection across all members of the combined arms teams was unexpectedly high, while the total investment in protection unexpectedly low. We inferred that, when agents were permitted to take on special roles and functions within the combined arms teams, detection was the most universally important attribute. It was also demonstrated that specialized combined arms teams could defeat combined arms teams with non-specialized members even when those teams were superior in capability. The usefulness or benefit imparted by mobility could not be determined at either level.

Finally, a games theoretic analysis of combined arms teams was conducted to explore the impact, costs and benefits of various candidate combined arms teams. This also provided a quantitative valuation of the consequences of adding, removing and replacing capabilities or assets within those teams. It was determined that there may not be a best combined arms team for the Land force. Combined arms teams may need to be self-adapting to defeat a range of different opponents. Thus, the potential benefits of agile and rapidly re-configurable forces are demonstrated, as are the potential limitations of forces with a static structure. Value in capability then comes from both the number of alternative capability options a force possesses and also the rate at which those options can be utilized.

References

1. Baker T, Botting M, Berryman M, Ryan AJ, Grisogono A-M and Abbot D (2004). Adaptive battle agents. In: Ryan AJ (ed). Proceedings of the Inaugural Complex Adaptive Systems in Defence Workshop 2004. Defence Science and Technology Organisation, Edinburgh, Australia.
2. Commonwealth of Australia (2004). JP 5000 Networking the Land Battlespace—Implementing the Concept. Future Land Warfare Branch: Army Headquarters, Australia.
3. Curtis NJ and Dortmans PJ (2001). A semiquantitative construct to identify opportunities for technology insertion using a generic description of the land forces. In: Puri V, Filippidis D, and Cantwell J (eds). Proceedings of the Land Warfare Conference 2001. Defence Science and Technology Organisation, Edinburgh, Australia, pp 364–380.
4. Goldberg DE (1989). Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley: Reading, MA.
5. Holland JH (1975). Adaptation in Natural and Artificial Systems. University of Michigan Press: Michigan.
6. Lauren MK, Stephen RT and Anderson MA (2001). MANA Version 1.0 Users Manual. DTA Technical Note 1/4. Defence Technology Agency: Auckland, New Zealand.
7. Nash JF (1950). The bargaining problem. Econometrica 18: 155–162. | Article |
8. Nash JF (1951). Non-cooperative games. Ann Math 54: 286–295. | Article |
9. Saaty TL (1977). A scaling method for priorities in hierarchical structures. J Math Psychol 15(3): 57–68. | Article |
10. Saaty TL (1980). The Analytic Hierarchy Process. McGraw-Hill: New York.
11. Shine D (2005). An Exploratory Study of the Army-as-a-System Core Skills. J Battlefield Technol 8(1): 39–44.
12. Spears WM, De Jong KA, Baeck T, Fogel DB and Garis H de (1993). An overview of evolutionary computation. In: Brazdil PB (ed). Proceedings of the European Conference on Machine Learning (ECML-93). Springer Verlag, Berlin, pp 442–459.
13. Von Neumann J and Morgenstern O (1944). The Theory of Games and Economic Behaviour, 2nd edn. Princeton University Press: New York.
14. Wheeler S (2005). A Theoretic Analysis of Combined Arms Teaming. Technical Report, DSTO-TR-1748. Defence Science and Technology Organisation: Edinburgh, Australia.