Answer Set Grammar for Efficient Generation of Policies in Coalition Environments
Military / Coalition Issue
Coalition parties are governed by their own sets of policies for guiding their operations. Policy technologies can be used to manage IT systems and networks, but they rely on centralized services. Coalition environments are highly dynamic and with no access to centralized infrastructures. Systems have to autonomously adapt their policies to maximize their effectiveness while minimizing risk. Self-generative policies can provide a key solution to self-autonomous management of devices in coalition setting. But they have to be learned in a context-dependent manner to accommodate changes and resolve conflicts with policies from other parties’ devices.
Core idea and key achievements
Grammars characterise a language and provide mechanisms for automatically generating acceptable sentences. Policies can be seen as sentences of a given language, and the self-generative policy model, from which they are derived, as the grammar that defines “acceptable” policies. In coalition settings, acceptable policies need to not only conform to a syntactic form but also satisfy context-specific semantic properties to guarantee their enforceability. Grammars of self-generative policies need to be context-sensitive. We have proposed a new class of context-sensitive grammars, called Answer Set Grammars (ASG), as a self-generative policy model. We have defined a formal language to specify these models as production rules annotated with semantic constraints expressed in Answer Set Programming. These annotations capture context-specific semantic properties that need to take into account during the policy instantiation process. Policies derived by the ASG’s production rules and satisfying the context-dependent semantic constraints are acceptable policies. We have developed two algorithms for (i) learning ASG from examples of context-dependent decisions, and (ii) instantiating policies from a learnt ASG, respective.
We have applied the framework to autonomous vehicles and logistic resupply of coalition forces and shown how intelligent devices can learn generative policies from few examples of past decisions, achieving higher accuracy than existing machine learning systems for classification. These two applications demonstrate our generative policy model to be flexible, able to learn from few examples, to explain learned outcomes and to capture human-driven policy rules.
Implications for Defence
Military operators will be able to use our ASG algorithms on autonomous devices to enable them to learn and instantiate context-dependent optimal policies in response to changes in the environment in, and in the coalition partners within which they operate.
Readiness & alternative Defence uses
Our system is at TRL 4 level. It has been validated in the lab and is ready to be tested in a real environment. It has also been used in other DAIS achievements (e.g., 1c.02, 1c.07, 1c.16).
Resources and references
- Law, Mark, et al. “Representing and learning grammars in answer set programming.” Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. No. 01. 2019.
- Bertino, Elisa, et al. “Generative Policies for Coalition Systems-A Symbolic Learning Framework.” 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS). IEEE, 2019.
Organisations
Imperial College London, Purdue University and Universitat Pompeu Fabra.