Marihan Shehata, Mohammed Moness and Ahmed M. Mostafa, Minia University, Egypt

Hyperparameter tuning plays a critical role in reinforcement learning (RL), particularly in safety-critical domains such as autonomous driving. In this work, we conduct a large-scale empirical analysis of hyperparameter sensitivity for two of the most widely used RL — Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) — using the CommonRoad-RL framework and the highD dataset. Functional analysis of variance (FANOVA) is employed to quantify main and interaction effects. Results show that performance variation in both algorithms is dominated by hyperparameter interactions, accounting for over 90% in PPO and nearly 88% in SAC, contrasting prior findings in simpler RL benchmarks. PPO is most sensitive to value learning and gradient stability, whereas SAC is driven by replay and training parameters. These findings highlight the need for interaction-aware tuning strategies to ensure robust RL deployment in complex driving tasks.

Autonomous driving, Proximal Policy Optimization (PPO), Soft Actor-Critic (SAC) Hyperparameter optimization, Hyperparameter sensitivity