This article explores the integration of Automated Guided Vehicles (AGVs) in warehouse order picking, a crucial and cost-intensive aspect of warehouse operations. The booming AGV industry, accelerated by the COVID-19 pandemic, is witnessing widespread adoption due to its efficiency, reliability, and cost-effectiveness in automating warehouse tasks. Through the strategic use of AGVs, this article focuses on enhancing the picker-to-parts system, which involves workers travelling to item locations, collecting them, and moving to the next location. We propose a novel MDP model for coordinating a hybrid team of human and AGV workers, aiming to maximize order throughput and operational efficiency, and employ a Neural Approximate Dynamic Programming (NeurADP) approach as the solution method. Specifically, our solution framework involves innovative solutions for non-myopic decision making, order batching, and battery management. The numerical results demonstrate that the NeurADP policy outperforms all benchmark policies, including both myopic and non-myopic ones, with a 3.32% and 5.44% improvement in order fulfillment over the alternatives. Comprehensive empirical analysis offers valuable insights for managing a heterogeneous workforce in a hybrid warehouse setting, highlighting the contributions of our work to the field of warehouse automation and logistics.

In requirements engineering (RE), anaphoric ambiguity is a frequent cause of misunderstandings. It can have a detrimental effect on the quality of requirements and jeopardize the success of a project. If stakeholders of the system, such as testers, developers, or customers, have different understandings or interpretations of software requirements, the system may not be accepted during customer validation. Despite its significance, there has been limited investigation into anaphoric ambiguity in RE. However, focusing on both recognizing and solving uncertainty can be more advantageous than just identifying it. Therefore we investigated the effectiveness of various QA learning techniques including encoder-based and text generation-based NLP models for two goals. We conduct detailed numerical experiments using various transformer models on two public requirements datasets and one generic dataset. Our results indicated that our QA architecture exhibits superior performance compared to baseline models in detecting ambiguity as well as resolving anaphora in contrast to other baseline approaches. We showed that our developed architecture can automatically support requirement development to minimize interpretation risk between stakeholders.

Sector duration optimization (SDO) is a problem arising in treatment planning for stereotactic radiosurgery on Gamma Knife. Given a set of isocenter locations, SDO aims to select collimator size configurations and irradiation times thereof such that target tissues receive prescribed doses in a reasonable amount of treatment time and healthy tissues nearby are spared. We present a multiobjective linear programming model for SDO to generate a diverse collection of solutions so that clinicians can select the most appropriate treatment. We develop a generic two-phase solution strategy based on the ε-constraint method for solving multiobjective optimization models, 2phasε, which aims to systematically increase the number of high-quality solutions obtained, instead of conducting a traditional uniform search. To improve solution quality further and to accelerate the procedure, we incorporate some general and problem-specific enhancements. Moreover, we propose an alternative version of 2phasε, which makes use of machine learning tools to reduce the computational effort. In our computational study on eight previously treated real test cases, a significant portion of 2phasε solutions outperformed clinical results and those from a single-objective model from the literature. In addition to significant benefits of the algorithmic enhancements, our experiments illustrate the usefulness of machine learning strategies to reduce the overall run times nearly by half while maintaining or besting the clinical practice. History: Accepted by Paul Brooks, Area Editor for Applications in Biology, Medicine, and Healthcare. Funding: This work was supported in part by the Natural Sciences and Engineering Research Council of Canada [Discovery Grant RGPIN-2019-05588]. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplementary Information [ https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2022.1252 ] or is available from the IJOC GitHub software repository ( https://github.com/INFORMSJoC ) at [ http://dx.doi.org/10.5281/zenodo.7048848 ].

Bug triaging is a critical task in any software development project. It entails triagers going over a list of open bugs, deciding whether each is required to be addressed, and, if so, which developer should fix it. However, the manual bug assignment in Issue Tracking Systems (ITS) offers only a limited solution and might easily fail when triagers are required to handle a large number of bug reports. During the automated assignment, there are multiple sources of uncertainties in the ITS, which should be addressed meticulously. In this study, we develop a Markov decision process (MDP) model for an online bug triage problem. In addition to an optimization-based myopic technique, we provide an ADP-based bug triage solution, called ADPTriage, which has the ability to reflect the downstream uncertainty in the bug arrivals and developers’ timetables. Specifically, without placing any limits on the underlying stochastic process, this technique enables real-time decision-making on bug assignments while taking into consideration developers’ expertise, bug type, and bug fixing time. Our result shows a significant improvement over the myopic approach in terms of assignment accuracy and fixing time. We also demonstrate the empirical convergence of the model and conduct sensitivity analysis with various model parameters. Accordingly, this work constitutes a significant step forward in addressing the uncertainty in bug triage.

Abstract Partially observable Markov decision processes (POMDPs) are a well‐established sequential decision making framework. Once a problem is modeled using this framework, a suitable POMDP solution algorithm is employed to obtain a policy that guides the user throughout the decision making process. However, POMDPs are notoriously difficult to solve to optimality. Therefore, there exist many approximate solution algorithms that are designed to generate policies for large‐scale POMDP models. On the other hand, many such approaches lack performance guarantees in terms of the solution quality. In this article, we focus on exact solution methods as well as approximate methods that provide bounds on the optimal value. Specifically, we investigate the performance improvements for the POMDP solution algorithms obtained through distributed implementations. We provide a detailed empirical analysis on various test problems, which highlights the benefits of the proposed approach.

Abstract In this study, we compare the Bayesian personalized ranking (BPR) algorithms with two recent state‐of‐the‐art algorithms, namely, noisy‐label robust Bayesian point‐wise optimization (NBPO) and Light Graph Convolution Network (LightGCN) algorithms, to validate and generalize their performance by using six publicly available datasets and one proprietary dataset containing web‐based data visualization usage records. We follow the guidelines explained in the original studies to pre‐process the input data and evaluate these algorithms using various evaluation metrics. We also perform hyperparameter tuning for the recommendation algorithms to determine the optimal configuration resulting in the best recommendation quality. We observe that the best hyperparameter configuration varies based on the algorithms and the datasets. The results of our analysis show some similarities with the results of the original studies while differing in certain respects. We observe that adaptive oversampling BPR (AOBPR) and LightGCN algorithms generate higher quality recommendations than the other algorithms. However, algorithm convergence rates vary significantly for each dataset. We note that the AOBPR approach is particularly useful for data visualization recommendation task, and can contribute to the improved recommendations in practice.

In software engineering practice, fixing a bug promptly reduces the associated costs. On the other hand, the manual bug fixing process can be time-consuming, cumbersome, and error-prone. In this work, we introduce a bug triaging method, called Dependency-aware Bug Triaging (DABT), which leverages natural language processing and integer programming to assign bugs to appropriate developers. Unlike previous works that mainly focus on one aspect of the bug reports, DABT considers the textual information, cost associated with each bug, and dependency among them. Therefore, this comprehensive formulation covers the most important aspect of the previous works while considering the blocking effect of the bugs. We report the performance of the algorithm on three open-source software systems, i.e., EclipseJDT, LibreOffice, and Mozilla. Our result shows that DABT is able to reduce the number of overdue bugs up to 12%. It also decreases the average fixing time of the bugs by half. Moreover, it reduces the complexity of the bug dependency graph by prioritizing blocking bugs.

Forecasting stock price movements is an important task for investors and traders, though still very difficult due to the unstable nature and complex behavior of the stock market. Accordingly, each piece of information related to stock prices can be deemed useful. In this regard, social media platforms provide a vast amount of information that can be used to predict stock movements. In this study, we compare the performances of various traditional, deep learning, and state-of-art pre-trained transformer models for text classification of tweets related to the stock market, which are obtained through a financial microblog, StockTwits. For this purpose, we collected 100,000 labeled messages of five stocks, namely, Apple Inc. (AAPL), Amazon (AMZN), Boeing Co, (BA), Walt Disney Co. (DIS), and the SPDR S&P 500 ETF Trust (SPY) for a period between December 2019 and June 2020. We used logistic regression and random forest as traditional classifiers and Long Short Term Memory and Gated Recurrent Unit as the deep learning algorithms, and BERT, DistillBERT, RoBERTa, and XLNet as the state-of-art transformer models to classify tweets as either “bearish” or “bullish”. Our numerical study showed that RoBERTa outperformed traditional classifiers and deep learning algorithms in terms of average F1-scores.

AIM: Having a reliable prediction model of radio signal strength is an essential tool for planning and designing a radio network. Given a geographic region, and associated power estimates linked to the transmitter placements, our objective is to develop machine learning models to predict the strength of the radio signals. BACKGROUND: The propagation model is often used to determine the optimal location of radio transmitters in order to optimize the power coverage in a geographic area of interest. However, it is often a costly operation to obtain the exact power measurements over a region for a given set of transmitter locations. Therefore, fast prediction methods are needed to estimate the power values given limited data. METHODOLOGY: We consider a dataset consisting of simulated power at each point in an environment for a given set of transmitter locations. We experiment with various machine learning models, namely, generalized linear models (GLMs), neural networks (NNs), and k-nearest neighbor (KNN), to estimate the power values for a given transmitter placement. We investigate various feature engineering approaches to enhance the predictive performance of the machine learning models. RESULTS: We observe that employed feature engineering methods such as polynomial degrees and transmitter to cluster distances significantly improve the prediction accuracy. In particular, GLM model performance notably improves thanks to these extracted features, where mean absolute error (MAE) is reduced around 77% from 11.37 dB to 2.55 dB. We note that KNN with k = 2 and DNN models perform better than NN and GLM. KNN has the best performance with an average MAE of 0.65dB and also substantially faster to train than NN/DNN models. In addition, our analysis shows that, to train a well-performing machine learning model, it is sufficient to use a dataset consisting of measurements at a fraction of the potential transmitter locations in a given region. CONCLUSIONS: Machine learning methods are highly effective for the coverage prediction task. Using carefully engineered features, simple models such as GLMs and KNNs can be as effective as more complex ones, especially for small datasets.