Reinforcement Learning with Adaptive Discount Factor for Clutch Judder Suppression with Stable Learning in Two-Speed EV Transmission

Designing the gear shift control for an automotive transmission is a complex task because it involves handling nonlinear behaviors like changes in friction between clutch plates and fluctuations in oil temperature. While deep reinforcement learning (DRL) has recently been used to reduce shift shock, most existing methods don’t account for real-world changes such as transmission aging. One major issue that becomes worse with aging is clutch judder—a type of vibration caused by wear. Traditional reinforcement learning assumes that the environment stays the same, which can lead to unstable learning when conditions change, making it hard to consistently reduce shift shock. To address this, we propose a new algorithm that adapts to aging transmissions by adjusting the discount factor—a key parameter in reinforcement learning that balances short-term and long-term rewards. Instead of keeping this factor fixed, our method starts with a lower value to ensure stable learning and gradually increases it to improve long-term performance. We use the loss function, which reflects how well the model is learning, as a signal to control the discount factor using a PID controller. To make tuning easier, we apply a model matching approach to set the PID parameters. Simulation results show that this method not only reduces shift shock more effectively but also keeps the learning process more stable compared to traditional fixed-discount approaches.