Optimizing Demand Forecasting: A Framework With Bayesian Optimization Embedded Reinforcement Learning for Combined Algorithm Selection and Hyperparameter Optimization

Demand forecasting plays an important role in various fields, and machine learning (ML) has emerged as a prevailing method to perform this task. The selection of an appropriate machine learning algorithm and hyperparameter optimization is essential for accurate predictions. However, this process can be complex and computationally demanding. In this paper, we propose a new framework to effectively tackle the Combined Algorithm Selection and Hyperparameter Optimization (CASH) problem. Our method utilizes reinforcement learning (RL) for ML algorithm selection and Bayesian optimization (BO) for hyperparameter tuning. The framework integrates a carefully designed reward function with an ε-greedy policy to guide the system in discovering the best ML pipeline and hyperparameter set. We extensively test the framework on small and large demand forecasting datasets, and the experimental results verify its ability in achieving high forecasting accuracy while significantly reducing computational time.