Hyperparameter Optimization Using Budget-Constrained BOHB for Traffic Forecasting

Machine learning models learn the model parameters from data but need the hyperparameters to be set by the model users. Hyperparameters are additional parameters specified during the machine learning model or algorithm’s training phase and substantially impact the model’s performance. Some popular hyperparameter tuning algorithms are grid search, random search, Hyperband, Bayesian optimization, and Bayesian Optimization Hyperband (BOHB). The selected search space and navigation approach affects how well the algorithms perform. The existing algorithms are either explorative in trying to find new configurations and are computationally intensive or exploitative in assessing the configurations but missing out on the optimum. In this work, we present an efficient hyperparameter optimization algorithm called Budget-Constrained Bayesian Optimization Hyperband (BCBOHB) to find an optimal set of hyperparameters in a limited number of iterations without compromising accuracy. The optimal configuration of hyperparameters is then tested on a traffic forecast model that uses the Long Short-Term Memory (LSTM) network. The findings demonstrate that the BCBOHB method is more accurate and can locate the ideal configuration in a constrained budget compared to contemporary techniques.