DEVS-NN: Efficient Development of Data Driven Models through Hybrid DEVS/SES/Neural Network Methodology

DEVS-NNtakes two complementary approaches to reduce the amount of data that must be collected for artificial intelligence/machine learning (AI/ML): 1. Use existing knowledge (in the form of a Discrete Event System Specification (DEVS) simulation model created by subject matter experts) to reduce the “work” a neural network (NN) has to do. The DEVS simulation model gets close to the correct output. The neural network needs less training data because instead of creating output from scratch it can use the DEVS simulation model output as a starting point. 2. A DEVS ontology (provided by System Entity Structures (SES)) is used to guide sampling/data collection. As we will detail later, a further advantage of our approach is that the SES ontology can provide additional validation benefits over unguided sampling/data collection approaches. Intelligently guiding the sampling/data collection process means less data needs to be collected. Our approach is most applicable to data collection from hardware-in-the-loop and high-fidelity physics simulations. In addition to being expensive, both of these data sources allow for a large degree of control over the scenarios that are simulated. The SES ontology will organize the set of possible scenarios; also called the scenario space. Both data sources have large scenario spaces, which will benefit from being organized using an SES ontology. Both data sources can be efficiently simulated using DEVS simulation models. The scenario space can be defined using SES, then programmatically pruned into a Pruned Entity Structure (PES), which can subsequently be automatically converted into a runnable DEVS simulation model. All of this can be implemented within RTSync’s proprietary MS4 Me DEVS Integrated Development Environment (IDE). Approved for Public Release | 21-MDA-11013 (19 Nov 21)