Resource-Aware Predictive Models in Cyber-Physical Systems
Cyber-Physical Systems (CPS) are composed of computing devices interacting with physical systems. Model-based design is a powerful methodology in CPS design in the implementation of control systems. For instance, Model Predictive Control (MPC) is typically implemented in CPS applications, e.g., in path tracking of autonomous vehicles. MPC deploys a model to estimate the behavior of the physical system at future time instants for a specific time horizon. Ordinary Differential Equations (ODE) are the most commonly used models to emulate the behavior of continuous-time (non-)linear dynamical systems. A complex physical model may comprise thousands of ODEs that pose scalability, performance and power consumption challenges. One approach to address these model complexity challenges are frameworks that automate the development of model-to-model transformation.
In this dissertation, a state-based model with tunable parameters is proposed to operate as a reconfigurable predictive model of the physical system. Moreover, we propose a run-time switching algorithm that selects the best model using machine learning. We employed a metric that formulates the trade-off between the error and computational savings due to model reduction.
Building statistical models are constrained to having expert knowledge and an actual understanding of the modeled phenomenon or process. Also, statistical models may not produce solutions that are as robust in a real-world context as factors outside the model, like disruptions would not be taken into account. Machine learning models have emerged as a solution to account for the dynamic behavior of the environment and automate intelligence acquisition and refinement. Neural networks are machine learning models, well-known to have the ability to learn linear and nonlinear relations between input and output variables without prior knowledge. However, the ability to efficiently exploit resource-hungry neural networks in embedded resource-bound settings is a major challenge.
Here, we proposed Priority Neuron Network (PNN), a resource-aware neural networks model that can be reconfigured into smaller sub-networks at runtime. This approach enables a trade-off between the model's computation time and accuracy based on available resources. The PNN model is memory efficient since it stores only one set of parameters to account for various sub-network sizes. We propose a training algorithm that applies regularization techniques to constrain the activation value of neurons and assigns a priority to each one. We consider the neuron's ordinal number as our priority criteria in that the priority of the neuron is inversely proportional to its ordinal number in the layer. This imposes a relatively sorted order on the activation values. We conduct experiments to employ our PNN as the predictive model in a CPS application. We can see that not only our technique will resolve the memory overhead of DNN architectures but it also reduces the computation overhead for the training process substantially. The training time is a critical matter especially in embedded systems where many NN models are trained on the fly.