Editorial: Special issue on operations research and machine learning

Preface

Many machine learning techniques work through optimizing specific objective functions. Supervised learning techniques are to minimize the prediction error such as mean square error (MSE) and misclassification rate, or maximize the conditional likelihood, posterior probability, etc. Unsupervised learning techniques usually group instances into clusters in a way that instances within each group are optimally similar while they are distant from instances in other groups. In reinforcement learning, the goal of an agent is to maximize its cumulative reward. However, there is still room to exploit optimization and operations research (OR) in machine learning, and vice versa.

Both machine learning and OR can gain advantages through integration and interaction. Optimization and OR techniques play a pivotal role in mitigating machine learning challenges. From feature selection to handling incomplete data and imbalance learning, they can enhance model accuracy and diversity. Their versatile applications encompass addressing bias, selecting optimal training sets, and developing classifiers that minimize misclassification errors across classes, offering comprehensive solutions to multiple machine learning hurdles. Every machine learning technique has several hyper-parameters that should be tuned to select the model that achieves the best performance on the learning task at hand. Normally, there are multiple criteria (or objectives) such as bias, variance, complexity, level of explainability, and fairness to be considered in model selection. The existing approach to address multiple criteria in machine learning is to transform the problem into a single-objective optimization problem using, for example, a weighted sum approach. However, this bears the issue of setting the importance of objectives in one way or the other, which is not a straightforward task, nor does the single solution to a multi-objective problem provide insights into the trade-offs between the objectives being optimized. Multi-objective optimization and multi-criteria decision-making as OR techniques can provide an opportunity to meet these criteria in machine learning.

On the other hand, machine learning techniques can contribute to finding the optimal solutions and making the best decision efficiently. Machine learning techniques can automate the process of the problem reduction in combinatorial optimization. Unsupervised learning can be used in Pareto pruning methods for multi- (or many-) objective optimization problems. Supervised learning can guide solutions through iterations to achieve convergence faster. Reinforcement learning can learn optimal controllers during the optimization process to improve the performance of optimizers. We are pleased to publish this special issue on ‘Operations Research and Machine Learning’ consisting of five articles.

In ‘Real-Time Production Scheduling Using a Deep Reinforcement Learning-Based Multi-Agent Approach’, Namoura et al. introduce a novel Deep Reinforcement Learning-Based Multi-Agent (DRLBMA) approach for real-time scheduling in flexible manufacturing systems, integrating multiple dispatching rules and employing a two-level self-organizing map to determine system states. The proposed DRLBMA approach is compared with other scheduling strategies and is found to be more efficient in terms of total weighted tardiness, throughput, and mean cycle time performance criteria. The study discusses potential future directions, including testing different rewarding policies, performing sensitivity analysis, and applying optimization techniques to enhance the learning process, while also suggesting research on deep neural networks for improved convergence and reduced computing costs, consideration of uncertain parameters, and scalability for general flexible manufacturing systems.

In ‘A reinforcement learning based dynamic room pricing model for hotel industry’, Tuncay et al. introduce a reinforcement learning-based dynamic room pricing model for the hotel industry to optimize room prices, maximize occupancy rates, and increase revenue. The model is trained using sales data from multiple hotels and employs the Q-learning algorithm with a unique reward function that considers profit and demand. It addresses the cold-start problem for new hotels by leveraging information from similar hotels, demonstrating improved performance with less error on test data. The study emphasizes the potential benefits of the model for tourism agencies, hotels, local economies, and policymakers in making efficient pricing decisions, improving customer experience, and fostering sustainable growth in the tourism industry. However, implementing the model may involve technical complexities, data synchronization challenges, and the need for stakeholder involvement. Future research aims to enhance the model by integrating additional features and conducting real-world tests.

In ‘The Recommender Problem with Convex Hulls’, Bougnol and Dula introduce a new method that emphasizes user-based collaborative filtering in the recommender problem, with a focus on selecting neighbours and weighting their ratings to make accurate predictions. This method uses computational geometry, specifically, convex hulls and linear programming (LP) to identify a target user’s neighbours and make personalized recommendations. The paper compares the effectiveness of this new method with Pearson correlations on three data sets: Jester, Yahoo! Music User Ratings of Musical Artists and Movielens. The results suggest that the proposed method provides more accurate predictive ratings, particularly on data sets of high density. This approach offers a new direction for addressing problems in machine learning using polyhedral geometry.

In ‘Cloud model-based best-worst method for group decision making under uncertainty’, Minaei et al. propose a multi-criteria group decision-making (MCGDM) under uncertainty using OR techniques as the best-worst method (BWM), cloud models and technique for order of preference by similarity to the ideal solution (TOPSIS). Data collection occurs in pairwise comparison via BWM to rank alternatives by the proposed MCGDM. The proposed method is applied to the selection of an online learning platform during the COVID-19 pandemic. The outcomes demonstrate improved conformity to decision-makers’ judgments over BWM and Bayesian BWM.

In ‘Warranty operation enhancement through social media knowledge: a deep-learning method’, Sarmast et al. propose a model that uses multi-channel social media data to identify common product defects and optimize warranty services. The approach includes the use of data mining, text mining, sentiment analysis, and deep learning techniques to analyse customer feedback on social media. Through this analysis, the model can identify potential issues with a product before warranty claims occur. The authors validate their model using data from Lenovo’s warranty service and claim the system can be implemented across multiple industries to improve product quality and customer satisfaction.

As Guest Editors of this special issue, we thank the Editor-in-Chief, Associate Prof. Joe Naoum-Sawaya, and the editorial office of INFOR: Information Systems and Operational Research for their support of our special issue. We also would like to express our sincere appreciation to the reviewers for their valuable comments.

Hadi A. Khorshidi
Melbourne School of Population and Global Health, the University of Melbourne, Melbourne, Victoria, Australia
[email protected]

Marzieh Soltanolkottabi
Mechanical and Industrial Engineering Department, University of New Haven, West Haven, CT, United States
[email protected]

Richard Allmendinger
Alliance Manchester Business School, the University of Manchester, Manchester, UK
[email protected]

Uwe Aickelin
School of Computing and Information Systems, the University of Melbourne, Melbourne, Victoria, Australia
[email protected]

Disclosure statement

No potential conflict of interest was reported by the author(s).