SCRO / Journées de l'optimisation

HEC Montréal, 29-31 mai 2023


HEC Montréal, 29 — 31 mai 2023

Horaire Auteurs Mon horaire

ABML Application based on Machine Learning

30 mai 2023 15h30 – 17h10

Salle: Société canadienne des postes (jaune)

Présidée par Ashkan Amirnia

4 présentations

  • 15h30 - 15h55

    Understanding Animal Foraging Behaviour Using Inverse Reinforcement Learning

    • Litong Zheng, prés., University of Toronto
    • Yingcong Tan, University of Toronto
    • J. Christopher Beck, University of Toronto
    • T. Jean M. Arseneau-Robar, University of Toronto
    • Julie A. Teichroeb, University of Toronto

    Understanding animal foraging behaviour is crucial to the study of behavioural ecology. However, the foraging decision-making process is complex as it depends on various factors such as the animal’s preference and resource quality. In this study, we modelled the foraging behaviour as a Markov decision process (MDP) with a linear reward function, and we adapted the maximum entropy inverse reinforcement learning (MaxEnt IRL) framework to predict the behaviour of wild vervet monkeys (Chlorocebus pygerythrus). Given the monkey’s foraging trajectories, the algorithm learns the parameters relevant to its decision-making process. Notably, the algorithm extends MaxEnt IRL to learn not only the weights in the reward function, but also the discount factor. Learning the former reveals the monkey’s preferences, while the latter represents how much it values future rewards. We applied the algorithm to two scenarios: foraging alone or in competition. We showed that, by using simple features, our approach can not only predict the most probable path the monkey would take, but also recover the distribution of possible paths, while previous work only focused on the monkey’s first and second decision. Our proposed methodology provides a new way to better understand foraging behaviour and can guide future experimental design.

  • 15h55 - 16h20

    Predicting consumer engagement using facial features extracted from movie trailer videos

    • Nadia Maarfavi, prés., Ontario Tech University
    • Salma Karray, Ontario Tech University

    Movie studios employ trailers as a marketing tool to generate interest and increase theater attendance. As a result, it is crucial to assess the efficacy of trailers before the movie's release to plan effective marketing strategies. One way to investigate the efficacy is by studying viewers' engagement. Previous studies have shown that movie information such as genre and budget can predict engagement towards trailers. This paper investigates for the first time the value of using information about the trailer content, namely those obtained from the actors' faces, in predicting engagement towards the trailer. Our goal is to address the following question: When predicting viewers' engagement toward a movie trailer, does adding the actors' facial attributes, such as their age, race, emotion, and gender, improve the predictive performance? If so, what facial attributes are most important? We collected movie information from IMDb and used the DeepFace library for facial attribute extraction and different machine-learning models for prediction. Results showed that the KernelRidge model performed best, and adding the actors' facial attributes improved the prediction accuracy by 7.34% for comments and 4.95% for viewers' sentiments. The most important facial features were "FaceNo," "RatioFaceNo," "Male," and "AverageAge."

  • 16h20 - 16h45

    Machine Learning Approach for Spare Part Inventory Management of Airport Baggage Handling System (BHS)

    • Yonghyuk Shin, prés., Korea University
    • Chulung Lee, Korea University

    Airport baggage handling system is becoming increasingly important along with the increasing demand for flights but, research on how to manage inventory according to the characteristics of the airport baggage handling system has not been carried out.
    The key elements of inventory management are divided into three aspects: 'inventory classification,' 'demand forecasting,' and 'inventory control policy.' Currently, various inventory classification methods and demand forecasting methods using machine learning are being studied, and inventory management policies that can be applied to various conditions are also being developed. However, in order to implement the most efficient inventory management, it is necessary to determine the optimal combination of the key elements of inventory management beyond the improvement of each methodology.
    This study aims to derive an optimal combination that reduces the total inventory cost of airport baggage handling systems by utilizing machine learning techniques for classification/forecasting and distribution-free models. This study is analyzed through empirical data from an international airport and is expected to contribute to the improvement and advancement of the airport baggage handling system.

  • 16h45 - 17h10

    Developing a multi-agent reinforcement learning-based model for disassembly planning with collaborative robots

    • Ashkan Amirnia, prés., Polytechnique Montreal
    • Samira Keivanpour, Polytechnique Montréal

    This research introduces a human-robot collaboration disassembly planning framework based on a multi-agent reinforcement learning model. A disassembly plan includes optimizing sequences to minimize operation time, energy consumption, and costs. By combining different optimization models and data-driven decision-making tools, it aims to handle not only the complexity but uncertainties of disassembly tasks. Few researchers have addressed practical issues such as real-time task allocation based on part features of end-of-life (EoL) products and learning-based algorithms in disassembly planning with collaborative robots (cobots). To tackle the mentioned gaps, this study proposes a multi-agent reinforcement learning-based model for disassembly planning. A graph embedding algorithm is proposed to vectorize the graph-based structure of an EoL product. Real-time disassembly sequence planning is also provided to consider the uncertainties in the operational context.