Learning Red Agent Policy from Observations for Neurosymbolic Autonomous Cyber Agents

The article presents a novel Policy Learning Technique that leverages imitation learning to train reinforcement learning agents in partially observable environments, specifically for autonomous cyber-defense applications. By integrating this technique with neurosymbolic approaches, such as behavior trees and learning-enabled components, the proposed method enables agents to predict the actions of cyber attackers (red agents) from network observations, achieving high prediction accuracy in various simulated scenarios. This advancement is significant for practitioners as it enhances the capability of autonomous agents to adapt and respond to evolving cyber threats in real-time.