Safety robustness of deep reinforcement learning in multi-agent scenarios

Multi-agent systems are being used to manage traffic, assigning tasks, regulating ant colonies, and operating self-driving cars, and drones. These systems involve multiple agents, coordinating, communicating, and working together with their surroundings to achieve the highest possible total numerical reward. Deep Reinforcement Learning (DRL) approaches are used to address these multi-agent applications. In many scenarios, the use of agents raise challenges to safety and robustness. To address these issues, in this work, we develop a DRL based system in which multiple agents interact with the real-world environment and act collaboratively and cooperatively. In our proposed model, several agents collaborate with one another to complete tasks and maintain a safe state. To take actions cooperatively and collaboratively of agents in accordance with the safety robustness of policies, we apply DRL algorithms such as proximal policy optimization (PPO) and Trust Region Policy Optimization (TRPO) algorithms and DRL approaches. We apply Curriculum Learning (CL) for their better performance and training. A reward structure is also proposed in this study which help agents to maintain their safe state. Number of Steps, Control and the reward for different policies are analysed as performance matrix in this study. The results shows that the safe policy adapted in the proposed model perform comparably better than the other policies.