KAN-LSTM Enhanced Multi-Agent Advantage Actor-Critic Reinforcement Learning for Autonomous Ramp Merging

Autonomous ramp merging represents one of the most challenging scenarios for autonomous vehicles (AVs), re-quiring safe interaction with human-driven vehicles (HDVs) while maintaining traffic efficiency. This paper proposes a novel KL-MA2C framework, which integrates Kolmogorov-Arnold Networks (KAN) and Long Short-Term Memory (LSTM) into the Multi-Agent Advantage Actor-Critic (MA2C) algorithm. The proposed framework addresses key limitations in existing approaches to robust safety mechanisms. Our key contribution lies in the synergistic integration of KAN-LSTM within a multiagent reinforcement learning framework. KAN provides superior function approximation through adaptive spline-based activation functions, while LSTM captures temporal dependencies in human driving patterns. This combination is enhanced by an action inspector that enforces safety constraints and Model Predictive Control (MPC) that ensures smooth trajectory execution. Extensive simulations demonstrate significant safety and efficiency performance improvements over state-of-the-art methods. These results establish KL-MA2C as an effective solution for autonomous ramp merging, with potential implications for safe and efficient autonomous driving in complex multi-agent environments.