Within the International Technology Alliance for Distributed Analytics and Information Sciences (DAIS-ITA)111https://dais-ita.org/, fundamental research is being performed to realise the future battlespace envisioned by various military Doctrine papers [1, 2]. The DAIS-ITA programme focuses on two technical areas; (1) Dynamic Secure Coalition Information Infrastructuresand (2) Coalition Distributed Analytics and Situational Understanding . In the future battlespace, coalition systems and devices will be required to operate in challenging environments that impose certain constraints such as lack of or low bandwidth connectivity to backend services, rapidly changing environmental conditions and the requirement to abide by legal regulations and mission directives. Therefore, coalition systems and devices require the capability to adapt and evolve such that they can behave autonomously ‘at the edge’ in previously unseen contexts. Crucially, systems need to understand the bounds in which they can operate based on their own (and that of other systems) capability, constraints of the environment and safety requirements. Also, systems and devices with varying autonomous capability may be required to collaborate with humans and other coalition partners to achieve a shared coalition goal, thus outlining the requirement for explainable actions and decisions. Recent work within the DAIS-ITA programme has developed the notion of Generative Policy-based Models (GPMs) for use in coalition environments  to enable coalition systems to define their behaviour and adapt to new situations whilst conforming to a variety of constraints and regulations.
In order to evaluate our work and demonstrate it’s applicability to coalition environments, illustrative scenarios and accompanying datasets are required. In a military scenario, this is challenging due to the lack of publicly available datasets and mission-specific information. Current efforts to date utilize non-military scenarios and datasets, such as open source traffic cameras for congestion classification  and IoT related deployments for smart-home policy management . Also, recent work has outlined an approach for generating synthetic data based on a Connected and Autonomous Vehicle (CAV) scenario [6, 7] utilising an agile knowledge representation framework that supports human-machine conversations as well as reasoning and hypothesis testing.
This paper extends the high level DAIS-ITA scenario  and develops the knowledge representation approach for synthetic data generation by enabling the user to control the complexity of the dataset to support future evaluations as the fundamental research progresses. The paper is structured as follows. In Section II we discuss a set of illustrative vignettes to motivate our work alongside relevant background knowledge. In Section III we outline the knowledge representation approach including DAIS-ITA specific domain models and discuss the approach for generating data with varying complexity. Finally, we conclude the paper in Section IV by providing closing remarks.
This section outlines a set of background knowledge including two military focused vignettes that may occur in a coalition operation, based on the high-level DAIS-ITA scenario . In the proposed vignettes, let us assume three partners are working together in a coalition, namely the UK, US and a fictitious Non-Government Organization ‘Kish’ in order to achieve a mission objective in a given region. In this section we also discuss a potential challenge for coalitions surrounding asset serviceability as a future application for generative policy research.
Ii-a Person of Interest Tracking
As detailed in , one possible mission the coalition could be tasked with is ‘Person of Interest Tracking’ which involves the monitoring of a High-Value Target (HVT). A detailed breakdown of the various phases of activity within this mission is shown in Figure 1 including potential adversary actions and resource constraints.
In this scenario, different coalition partners will have different roles and responsibilities. For example, UK forces could partner with Kish’s local security force to execute the ‘Plan’ and ‘Find’ stages and the US may partner with the UK and Kish for the ‘Track’, ‘Assess’ and ‘Exploit’ stages. The planning stage involves identifying a set of coalition assets and resources to use which may be constrained by power and connectivity, as well as being vulnerable to man-in-the-middle attacks from an adversary intercepting network communications. Then, the coalition finds and tracks the person of interest, and finally collates the results. Throughout the mission, coalition assets are subject to re-tasking, e.g. if the enemy plants an Improvised Explosive Device (IED), coalition forces and assets may have to relocate to deal with the IED. Also, the coalition has to manage data storage, power and network connectivity constraints dynamically throughout the mission.
Ii-B Logistical Resupply
The second vignette describes a logistical resupply mission to serve coalition troops that are stationed in a given location such as a city. A convoy is provided by a civilian contractor and is accompanied by a guarding force, or force protection element, from another coalition partner. In addition, regular resupply of troops in the city is conducted by autonomous resupply drones, that also have other tasks such as resupply of high priority special forces elements in mountainous locations outside the city. Similarly to the first vignette, this vignette is broken down into phases of activity as shown in Figure 2. These take into account elements such as the planning phase where intelligence, surveillance and reconnaissance assets will have to be assigned to the mission; monitoring of the mission during its execution; and the recovery phase back to base. The enemy will once again be employing disruption tactics and the coalition is constrained by the resources available for use.
Ii-C Military Environments
As detailed in , the missions outlined above could potentially be executed in a wide range of environments, which impose various constraints on the types of activities that can occur for a given mission. This is detailed in Table I.
|Number of non-combatants||High||Low||Low||Low|
|Amount of valuable infrastructure||High||Low||Low||Low|
|Presence of multi-dimensional battlespace||Yes||No||Some||Yes|
|Restricted rules of engagement||Yes||No||Some||Yes|
|Detection, observation, engagement ranges||Short||Long||Short||Medium|
|Avenues of approach||Many||Many||Few||Few|
|Freedom of movement & manoeuvre||Low||High||Low||Medium|
Ii-D ALFUS Framework: Autonomy Levels for Unmanned Systems
Huang has outlined varying levels of autonomy, ranging from – inclusive that describe the full range of autonomous capability in unmanned systems . Figure 3 details Mission Complexity (MC), Environmental Complexity (EC) and Human Interaction (HI) capabilities at each level. The overall ALFUS level is obtained through scoring each capability with a value – inclusive, for levels –. At level 0, each capability receives a 0 score to indicate no autonomous capability and level 10 remains independent to represent the ultimate goal for autonomy.
Ii-E Asset Serviceability
, coalition partners are required to work together which may involve sharing assets with one another. An asset could be a physical device such as a surveillance camera, an autonomous system such as an unmanned aerial vehicle or an asset could be a virtual service such as a database or a specific type of machine learning model such as a face recognizer or weaponry detector. Each asset has a set of time-bound constraints that govern it’s use at a given point in time. Possible constraint types are defined as follows:(1) Physical Constraints—the asset can only operate in certain physical locations or contexts. Also, in order to utilize an asset the requestor may need to be within a suitable physical range of the asset; (2) Security Constraints—certain functionality within the asset may be withheld by the asset owner through access control. For example, the high value targets database may only accept read requests from a requestor iff they are requesting information about a given target. (3) Trust Constraints—certain assets or functionality may only be shared between coalition partners iff the trust value between partners is above a predefined threshold
. In the above definitions, constraints can be local to the individual asset in a given context (e.g. a certain autonomous vehicle is currently unable to drive autonomously given high wind speeds) or constraints can be applied at the global level by the coalition partner (e.g. do not share the facial recognition model unless the trust with the requesting partner is greater than or equal to a predefined threshold). Given the constantly evolving context and potentially dynamic asset pool caused by coalition partners frequently bringing assets online/offline during a mission, local and global serviceability policies that define the type of requests that an asset or coalition partner can service will be constantly in flux and it won’t be possible for a human operator to generate all of these policies for all assets in all possible contexts in a timely manner. Also, given the distributed nature of a partner’s assets, the coalition partner as a whole requires means to keep track of the serviceability of its asset pool through a generated policy and also is responsible for communicating this updated policy in a distributed manner to its assets. For example, a US autonomous vehicle, denotedVehicle A moves to a new location in the city so that the camera and object detector on-board the vehicle can now fulfill requests for information regarding a new type of object (e.g. civilian pedestrians are now in view). Vehicle A now needs to generate a new local policy that states that it can service these new requests and this policy needs to be communicated to all US assets, so that the US can generate a global policy that states that if any asset within the US’ inventory receives a request regarding the street Vehicle A is located at, the asset can direct the requestor to Vehicle A, or, utilize the Software Defined Network to perform the request to Vehicle A on behalf of the requestor.
Utilizing the existing knowledge and literature presented in Section II, as well as the requirement to account for asset serviceability, in this section we discuss our approach for representing the coalition environment by creating DAIS-ITA specific domain models to support synthetic data generation.
Iii-a Knowledge Representation
ITA Controlled English (CE)  was chosen as a semantic knowledge base due to it’s expressive power and built-in reasoning capabilities that enable future human-machine conversational interfaces to be developed that utilize the conceptual model. Also, CE enables domain models to be easily specified using human-readable, controlled natural language which ensures our work is extendable if necessary. Figure 4 details the concept definitions for a mission and a mission instance in CE. This is inspired by the illustrative vignettes outlined in Figures 1 and 2 as missions contain high level stages, can be disrupted by potential adversary actions and are bound by certain constraints. Also, inspired by Table I, missions can be executed in different mission environments and different environmental conditions, as well as being undertaken by a certain coalition at a certain date/time. Note that the coalition, mission environment and environmental condition instance concepts are defined separately and are not included in Figure 4 due to space limitations. We release the full concept definition online for further reference222https://github.com/dais-ita/coalition-data.
The concept of a mission and a mission instance can then be instantiated in CE as facts and stored in the knowledge base. For example, Figure 5 defines the person of interest tracking mission and instantiates it with respect to a coalition, mission environment, environmental conditions and a start time. Domain modelling is repeated for other concepts including ALFUS levels, coalition partners, coalition trust relationships, environmental conditions and assets and stored in the knowledge base for instantiation. The full model illustrating concepts and their relationships is shown in Figure 6.
Iii-B Fact Generation
Facts such as those detailed in Figure 5 can be used to generate a dataset that is required for evaluating GPMs. However, in this scenario, there are a large amount of facts required in the knowledge base in order to represent a real-world coalition environment and it would be too cumbersome to generate these facts manually. Therefore, we define a small set of base concepts and generate facts programmatically as shown with the generalized approach in Algorithm 1.
Firstly, we define a set of possible environmental conditions: visibility level, temperature level, rainfall level, snowfall level, wind speed level and humidity level where each condition contains an associated lower and upper bound, appropriate units and a weighting to denote the importance of this condition. This refers to line 1 in Algorithm 1. Environmental condition values are then generated for each condition between the lower and upper bounds according to a user-configurable granularity parameter—i.e. if the humidity level has a lower bound of 0%, an upper bound of 100% and the granularity parameter is 5, values 0%, 25%, 50%, 75% and 100% would be generated. This refers to line 2 in Algorithm 1. To ensure coverage of a wide range of possible conditions, an environmental condition instance is generated for each possible combination of the environmental conditions—i.e. setting the granularity parameter to 5 and using the 6 different environmental conditions there would be different environmental condition instances which corresponds to line 3 in Algorithm 1. Also, during the combination process, we normalize each condition and compute a weighted average to denote the severity of each environmental condition instance, according to the importance weight specified for each condition. Then we iterate over all the possible combinations and build a CE fact sentence, such as those shown in Figure 5. This corresponds to lines – in Algorithm 1. Next, all the possible ALFUS scores are generated. Recall from Section II-D that an overall ALFUS score is the sum of the capability scores, which identifies the level of complexity of a mission an asset can deal with, the level of environmental complexity an asset can operate in and the amount of human involvement an asset requires. Given each capability can be scored between – inclusive and that there are 3 capabilities, we generate overall ALFUS scores, representing all possible combinations of scores for mission complexity, environment complexity and human interaction. We manually add an additional score for ALFUS level 10 as level 10 is independent from other ALFUS levels. Given all coalitions, mission environments, mission types, generated environmental condition instances and a predefined list of 4 start times, we generate all possible combinations of mission instances which ensures our dataset covers all possible conditions a mission could be executed under. Setting the granularity parameter for the environmental conditions to 5 and assuming there is 1 coalition, 4 mission environment types (Table I) and 2 mission types, we generate possible mission instances. To account for asset serviceability, we generate a set of assets of varying type including physical assets, autonomous assets and virtual assets. For the autonomous assets we utilize the generated ALFUS scores in order to specify varying levels of capability and constraints. We also specify an asset worth, to denote it’s value to a particular owner. For autonomous assets, the worth is set dynamically based on it’s autonomous capability. In order to represent the fluid notion of assets dynamically coming online/offline during a mission and to enable asset utilization requests to be modeled we define the concept of a live asset inventory that stores the real-time status of a collection of assets for every coalition partner on every mission instance. We then assign a random set of assets to each coalition partner’s inventory, where the number of assets to assign is user configurable. Also, we generate a random starting configuration for each asset including a latitude/longitude location, a risk of adversarial compromise and a Boolean value indicating if this asset is available to use. The bounding box used to generate the initial asset locations is also user configurable. Finally, we create the concept of an asset request which indicates that a coalition partner requested to use an asset from another coalition partner on a given mission instance at a particular time. The request is also annotated through user configurable logic to determine an approve or reject decision which outlines whether this request is successful or not.
Iii-C Customized Data Generation
This enables researchers to specify custom, potentially complex logic to define a policy model that can be used to evaluate future GPM architectures.
This paper has presented a conceptual domain model and dataset for use on the DAIS-ITA research programme. The domain model has been built using an agile representation framework which enables future work to extend our model easily using controlled natural language. Also, the dataset complexity can be controlled using custom logic declarations that will support future evaluations for GPMs
. Future work involves the ability to support partial data generation to simulate a real-world scenario where a complete dataset is not available and to investigate the suitability of Generative Adversarial Neural Networks for the suitability of synthetic data generation in the coalition environment.
This research was sponsored by the U.S. Army Research Laboratory and the U.K. Ministry of Defence under Agreement Number W911NF-16-3-0001. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Army Research Laboratory, the U.S. Government, the U.K. Ministry of Defence or the U.K. Government. The U.S. and U.K. Governments are authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation hereon.
- arm  “Stand-to!” Jun 2018. [Online]. Available: https://www.army.mil/standto/2018-06-06
- of Defence  M. of Defence, “Global strategic trends,” Oct 2018. [Online]. Available: https://www.gov.uk/government/publications/global-strategic-trends
- Verma et al.  D. Verma, S. Calo, S. Chakraborty, E. Bertino, C. Williams, J. Tucker, and B. Rivera, “Generative policy model for autonomic management,” in IEEE SmartWorld, Ubiquitous Intelligence Computing, Advanced Trusted Computed, Scalable Computing Communications, Cloud Big Data Computing, Internet of People and Smart City Innovation, 2017, pp. 1–6.
- Willis et al.  C. Willis, D. Harborne, R. Tomsett, and M. Alzantot, “A deep convolutional network for traffic congestion classification,” in Proc NATO IST-158/RSM-010 Specialists’ Meeting on Content Based Real-Time Analytics of Multi-Media Streams, 2017, pp. 1–11.
Goynugur et al. 
E. Goynugur, G. de Mel, M. Sensoy, K. Talamadupula, and S. Calo, “A knowledge
driven policy framework for internet of things,” in
Proceedings of the 9th International Conference on Agents and Artificial Intelligence, vol. 2, 2017, pp. 207–216.
- Cunnington et al.  D. Cunnington, G. de Mel, and D. Shaw, “Reduce cognitive burden on drivers through contextualising environments,” in 2018 IEEE 87th Vehicular Technology Conference (VTC Spring), June 2018, pp. 1–7.
- Cunnington et al.  D. Cunnington, I. Manotas, M. Law, G. de Mel, S. Calo, E. Bertino, and A. Russo, “An Answer Set Grammar-based Generative Policy Model for Connected and Autonomous Vehicles,” 2019, submitted for publication.
- White et al.  G. White, S. Pierson, B. Rivera, M. Touma, P. Sullivan, and D. Braines, “Dais-ita scenario,” in Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, vol. 11006. International Society for Optics and Photonics, 2019.
- Huang  H.-M. Huang, “Autonomy levels for unmanned systems (alfus) framework,” 2007.
- Braines et al.  D. Braines, D. Mott, S. Laws, G. de Mel, and T. Pham, “Controlled english to facilitate human/machine analytical processing,” in Next-Generation Analyst, vol. 8758. International Society for Optics and Photonics, 2013, p. 875808.