Suppressing Control Oscillations in Reinforcement Learning via External Action Filtering for Autonomous Driving
2025 16th International Conference on Electrical and Electronics Engineering, ELECO 2025, İstanbul, Türkiye, 27 - 29 Kasım 2025, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/eleco69582.2025.11329257
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Bursa Uludağ Üniversitesi Adresli: Evet
Özet
Deep Reinforcement Learning (DRL) agents operating in continuous action spaces for autonomous driving often exhibit high-frequency, non-human-like control oscillations, which compromise passenger comfort, predictability and social acceptance. This paper addresses this challenge by proposing a modular DRL architecture that decouples the primary control policy from the secondary objective of action smoothing. The framework features a Soft Actor-Critic (SAC) agent trained for a combined lane-keeping and static obstacle-avoidance task. The agent's raw outputs are processed by a dedicated, external filtering module before actuation. A systematic evaluation demonstrates that a Weighted Moving Average (WMA) filter reduces action volatility by over 92% compared to an unfiltered baseline, achieving exceptionally smooth, human-like control with only a minimal compromise in task accuracy. This result validates the decoupled approach as a robust and flexible paradigm for developing socially acceptable and comfortable autonomous driving systems.