Public opinion and persuasion of algorithmic fairness: assessment of communication protocol performance for use in simulation-based reinforcement learning training

Abstract

As the popularity of AI continues to grow, the techniques used to train AI systems have become increasingly intriguing. Reinforcement learning (RL) in simulation environments provides a safe and effective training approach for AI. However, the performance of the communication protocol between the reinforcement learning system and the simulation environment significantly impacts the overall system performance. This research aims to investigate the impact of communication protocols on algorithmic fairness by exploring public opinion and persuasion. The test software uses reinforcement learning data of floating-point numbers and images to assess communication methods, including Sockets, Socket.IO, gRPC, and ZeroMQ. Results indicate that Sockets and ZeroMQ exhibit similar performance in transferring floats, while ZeroMQ outperforms Sockets when it comes to transmitting images. Sockets are preferred when dealing with larger datasets. ZeroMQ stands out due to its speed and user-friendly simplicity, making it a suitable choice for reinforcement learning in simulation, employing Unreal Engine, AGX Dynamics, and Stable Baselines3. In the conducted experiments, reinforcement learning is found to be the most time-consuming component, followed by simulation, with communication constituting half of the total time. As the complexity of the system increases, the time spent on reinforcement learning and simulation is expected to grow faster than that of communication, emphasizing the need to optimize other components.