Targets in Reinforcement Learning to solve Stackelberg Security Games

TL;DR

This paper reviews how reinforcement learning can be applied to Stackelberg security games, emphasizing the importance of target representations for improving agent performance in security scenarios.

Contribution

It analyzes current RL modeling approaches for SSGs and explores potential enhancements in target representations to boost effectiveness.

Findings

Current RL models effectively simulate defender-attacker dynamics.

Improved target representations can enhance RL agent performance.

Potential for better security outcomes through refined modeling.

Abstract

Reinforcement Learning (RL) algorithms have been successfully applied to real world situations like illegal smuggling, poaching, deforestation, climate change, airport security, etc. These scenarios can be framed as Stackelberg security games (SSGs) where defenders and attackers compete to control target resources. The algorithm's competency is assessed by which agent is controlling the targets. This review investigates modeling of SSGs in RL with a focus on possible improvements of target representations in RL algorithms.