Understanding the road roughness profile is crucial for traffic management authorities to assess rider comfort and safety, as well as to understand its impact on bridge vibrations. Thus, accurately identifying the road roughness profile is an essential task. While direct methods like laser measurement devices are effective for determining road roughness profiles, they are costly because they require a specific measurement system. Therefore, identifying road roughness profiles indirectly through vehicle acceleration responses was investigated as it offers a faster and more cost-effective alternative to direct monitoring methods. This study aims to investigate the feasibility of enhancing the accuracy of road roughness profile identification using vehicle responses with Bayesian optimization. A method for directly identifying road roughness profiles from vehicle acceleration was proposed using Dynamic Regularized Least Squares (DRLS) minimization. This method improved both accuracy and computation efficacy compared to existing methods. Although this method has shown promising results, there are still challenges to address to improve identification accuracy further. These include certain limitations within the algorithm to reduce its robustness, particularly in tuning the regularization parameter which was tuned with L-curve method. The proposed method in this study applies Bayesian optimization to find the optimal regularization parameter in DRLS minimization. Moreover, a new regularization parameter that is adaptive to noise is proposed. The improvement results by adding Bayesian optimization to DRLS minimization is examined using measurement data from a bridge in Japan.