Authors:
Pravin Suryawanshi
1
;
Sandesh Gaikwad
1
;
Akansha Kumar
2
;
Akhil Patlolla
1
and
Sai Jayakumar
3
Affiliations:
1
Data Scientist, Jio Platforms Limited, Navi Mumbai, Maharashtra, India
;
2
Chief Data Scientist, Jio Platforms Limited, Hyderabad, Telangana, India
;
3
Product Manager, Jio Platforms Limited, Hyderabad, Telangana, India
Keyword(s):
Hyperparameter Optimization, Demand Forecasting, Genetic Algorithm, Bayesian Optimization, LSTM Network.
Abstract:
Demand forecasting is highly influenced by the non-linearity of time series data. Deep neural networks such as long short-term memory networks (LSTM) are considered better forecasters of such data. However, the LSTM network’s performances are subject to hyperparameter values. This study proposes a hybrid approach to determine the optimal set of hyperparameters of an LSTM model using Bayesian optimization and genetic algorithm. Bayesian optimization explores the search space in the direction where the improvement over the existing solution is likely, based on a fitness function. At the same time, a genetic algorithm is an evolutionary approach that can achieve global convergence by using selection, crossover, and mutation operators. The proposed hybrid approach utilizes the strengths of both these algorithms to tune the values of the hyperparameter of the LSTM network to minimize the forecasting error. In the dataset considered, we found that the hybrid approach reduced the forecastin
g error by approximately 27% compared to the Bayesian optimization approach. Additionally, the proposed method is better than the genetic algorithm when performed independently, with a decrease in error value by approximately 13%.
(More)