SUMMARY

Cyber-physical systems (CPS) integrate humans and computational systems to control physical plants using sensing signals and feedback loops. The interconnection, modularity, and exposure to human environments, reveal CPS to attacks from adversaries. In an adversarial, information-constrained, and model-agnostic environment, reinforcement learning (RL) methods design optimal control policies using reward signals that possess decentralized learning and deployment protocols. Increased continuous data-sharing between sensors, actuators, and the cloud adds complexities to the network leading to degradation of communication effectiveness and increased the possibility of reinforcement signal dropouts, leaving the learning framework with access to a sparse set of rewards.

This work proposes a Moving Target Defense (MTD) framework, to switch between allowable actuator combinations. MTD allows the input-generation framework to maintain controllability, enhance performance, and safeguard the CPS from. A data-driven controllability algorithm identifies allowable actuator combinations to guarantee controllability. During sensing breakdown, lack of complete reward signals dictates an intermittent framework to approximate value functions and tune embedded neural networks. When continuous communication between the reinforcement channel and the controller is challenged, the proposed framework will develop an intermittent learning algorithm that will provide conditions wherein the intermittent reinforcement signal is rich enough to allow learning of the optimal control policy.