Optimizing Deep Reinforcement Learning with Log Hyperbolic Cosine loss function: A Novel Approach

Abstract In a recent study [1], Huber loss function outperforms mean-square error (MSE) loss on the problem of balancing cart-pole. In this study, we introduce the Log Hyperbolic Cosine (Logcosh) loss and conduct a comprehensive comparison with the established alternatives. Fundamentally the presented work investigates the impact of loss function choice on the training performance of Deep Reinforcement Learning (DRL) algorithms. Our evaluation criteria encompass the achievement of average rewards equal to or exceeding 195, along with a frequency-based stability test. The experimental results reveal that Logcosh outperforms both MSE and Huber losses, showcasing a faster convergence and a substantial increase in rewards per episode. This significant performance enhancement suggests that Logcosh is a promising alternative for optimizing DRL control algorithms.