Agile Metamodel Inference using Domain-specific Ontological Languages

Modeling complex systems – for public health threats such as a pandemic like COVID-19 – is challenging because domain experts, data scientists, and programmers have limited capabilities to create actionable models that can be verified and modified in real-time. Currently, mathematical diagrams that represent complex systems are only informally related to the model; if data scientists make changes to diagrams, then domain experts must make manual changes to the model. This, in turn, leads to errors. 

There is also no simple way to tie results back to the diagrams or models.  This has posed a problem for responding to COVID 19, which has required agile planning by different agencies, as well as local, state, and federal government policymakers. 

DARPA’S Automating Scientific Knowledge Extraction program seeks to automate the processes involved with scientific discovery and application. Galois’s planned solution is the  Agile Metamodel Inference using Domain-specific Ontological Languages (AMIDOL) project. 

AMIDOL is designed to help scientists gain insight from models of complex systems by using the following: visual Domain Specific Languages (DSLs), an intermediate representation (IR), and a machine-assisted inference engine. AMIDOL is designed to be a customizable and flexible solution for policymakers and scientists managing pandemic response. 

 An Interface for Abstract Functions 

The DSLs are designed to simplify the modeling process by creating nouns and verbs to represent formal objects. For instance, using COVID 19 as an example, nouns would be created for “susceptible patient” and “infected patient.” They would be connected by verbs for infection and recovery. This allows scientists to work more efficiently since they wouldn’t need to write any code.  

The IR is designed to be state-level, providing notions of state variables and actions, as well as input/output predicates. Reward models implement the DSLs as classes of templates. Scientists would have the ability to test different hypotheses because all objects in a particular DSL have corresponding and implementable representations in the IR. This would be a vast improvement over the current modeling environment. 

Inferences Aided by Machine Learning

AMIDOL’s goal is to improve model clarity and explainability for scientists. The IR is being designed to define both rate and impulse reward models that provide a wide array of solutions for users. The inference engine will allow users to choose from a library of solvers. This solver-agnostic approach aims to give users flexibility in choosing metrics based on variables in the Intermediate Representation. 

AMIDOL also aims to help “clean” existing models that have errors in them.  When scientists import a model into AMIDOL, errors can be identified and often automatically repaired in the IR.  

Rapid Response Could Be “The New Normal”

For models developed within AMIDOL, the framework takes representations and automatically synthesizes code on the back-end. This ability to factor in (or factor out) data in the IR gives a scientist the ability of a software engineer without needing to actually be a software engineer.  Scientists could rapidly synthesize models and quickly verify if the data makes sense.

The COVID 19 crisis has provided a potential opportunity to make policy-based gains within governmental organizations. AMIDOL could particularly help policymakers respond to crises in real-time, making knowledge more actionable in the near-term. 

If another pandemic struck, AMIDOL could help domain experts, data scientists, and programmers create more accurate and updateable models in real-time, with far less ambiguity. 

Scientists could leverage existing scientific knowledge and not be hampered by a single implementation.  Every research paper could match the model – essentially ensuring reproducible science that other domain experts can use.

If you’d like to read about our most recent update with AMIDOL, please read our blog post here.

You can also view our project on GitHub here. 

Acknowledgment of Support and Disclaimer: The Performer shall include an acknowledgment of the Government’s support in the publication of any material based on or developed under this Agreement, stated in the following terms: “This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR00111990005.”