Deep Reinforcement Learning for Green Security Games with Real-Time Information

摘要: Green Security Games (GSGs) have been proposed and applied to optimize patrols conducted by law enforcement agencies in green security domains such as combating poaching, illegal logging overfishing. However, real-time information footprints agents' subsequent actions upon receiving the information, e.g., rangers following chase poacher, neglected previous work. To fill gap, we first propose a new game model GSG-I which augments GSGs with sequential movement vital element of information. Second, design novel deep reinforcement learning-based algorithm, DeDOL, compute patrolling strategy that adapts against best-responding attacker. DeDOL is built double oracle framework policy-space response oracle, solving restricted iteratively adding best strategies it through training Q-networks. Exploring structure, uses domain-specific heuristic initial constructs several local modes for efficient parallelized training. our knowledge, this attempt use Deep Q-Learning games.