Learning reserve prices in online repeated reverse auctions
Εκμάθηση τιμών επιφύλαξης σε άμεσες επαλαμβανόμενες αντίστροφες δημοπρασίες
View/
Keywords
Online machine learning ; Auctions ; Multiplicative weights update ; Follow the Perturbed Leader ; FTPL by Generically Leveraging Past Decisions ; FTPL-GLPD ; Submodular functions ; Online algorithmsAbstract
In this thesis, we are inspired from the work conducted by Roughgarden and Wang [1] to explore the concept of calculating non-anonymous reserve prices in repeated multiunit reverse auctions. More specifically, we address the problem of computing reserve prices in repeated reverse multiunit Vickrey auctions, focusing on two key scenarios: (i) determining non-anonymous reserve prices, where each bidder is assigned a personalized reserve price, and (ii) computing a single reserve price applicable to all bidders. For the first scenario, we introduce an offline algorithm that achieves a theoretical upper bound of 2 for the ratio of the optimum solution surplus over the algorithm’s surplus and extend it to an online setting using the Follow the Perturbed Leader (FTPL) algorithm. For the second scenario, we propose an offline algorithm that achieves n for the ratio of the optimum solution surplus over the algorithm’s surplus, with n being the number of bidders in the auction and adapt it into an online algorithm using the Multiplicative Weights Update (MWU) framework. Through extensive experimentation, we evaluate the algorithms' performance using metrics such as average daily surplus and regret. Bid data for these experiments are generated using autoregressive models to simulate realistic auction dynamics. Our results reveal that solutions with personalized reserve prices outperform those with a single reserve price in terms of surplus and regret. We also demonstrate the significant impact of parameter tuning on algorithm performance, emphasizing the versatility of our methods across diverse auction environments. Additionally, we develop FTPL-GLPD, a modified version of FTPL that incorporates historical data dynamically, achieving superior performance across most test cases. This work provides a robust framework for reserve price computation in online auctions, while drawing a comparison between a solution that computes non-anonymous reserve prices versus a solution that computes one global reserve price.
Postgraduate Studies Programme
Πληροφοριακά Συστήματα και ΥπηρεσίεςDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών ΣυστημάτωνNumber of pages
69Language
EnglishCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα