Framework introduction – Bachelor's theses

As is the custom at the HPI, each of the six original students that worked on the project wrote their respective bachelor's thesis on a topic related to the project work.

All of these theses offer in-depth information of a specific topic related to the project.

The Marketplace of the Future: Simulation of Market Processes in Re-Commerce

By: Nick Bessin (@SinNeax)
Language: German
German Title: Der Marktplatz der Zukunft: Simulation von Marktprozessen im Re-Commerce

Abstract:
to be inserted when complete

Links:
to be inserted when complete

Pricing in the Re-Commerce Domain: Analysis of Pricing Strategies with an Online Market Simulation

By: Leonard Dreeßen (@ldreessen)
Language: German
German Title: Preisfindung in der Recommerce-Domäne: Analyse von Preisstrategien mithilfe einer Online-Marktsimulation

Abstract:
to be inserted when complete

Links:
to be inserted when complete

A comparison of reinforcement learning algorithms for dynamic pricing in recommerce markets

By: Jan Niklas Groeneveld (@jannikgro)
Language: German
German Title: Ein Vergleich von Reinforcement-Learning-Algorithmen für die dynamische Preisgestaltung auf Recommerce-Märkten

Abstract:
Trading second-hand items on online marketplaces is a rising business. Thereby, pricing is a major challenge. E-commerce is highly dynamic, the product range is large, and fast reactions to competitors are required. Thus, automation of pricing strategies is a necessary step for traders. This thesis is about dynamic pricing with Reinforcement Learning (RL), a machine learning technology. Today, thanks to intensive research in the past years, many different RL-algorithms exist. Five of these algorithms are compared to competitive rule-based strategies in monopoly, duopoly, and oligopoly markets. It is found that A2C, PPO, and SAC can succeed and outperform competitors, while DDPG and TD3 lack the necessary performance. Later, several problems that arise in real-world applications are tackled and solved. It is shown that missing some information about the competitor does not break the algorithms. The problem of not knowing the opposite’s strategy can be successfully solved by self-play.

Links:
Download PDF

Scalable Learning in the Cloud

By: Judith Herrmann (@felix-20)
Language: German
German Title: Skalierbares Lernen in der Cloud

Abstract:
Cloud environments are becoming more and more important. Especially in the field of machine learning, they offer the possibility of making the learning process more efficient. The better hardware, the easy and continuous availability makes these remote computers indispensable for modern systems. Large providers often sell a complete solution that includes all the tools researchers need to work. However, only a few specialise in reinforcement learning. This paper presents a cloud architecture that can be used to train reinforcement learning agents. It is based on a market simulation in which the agents can learn the best possible pricing strategies. The requirements and the implementation are discussed. In addition, the scalability of the architecture will be shown on the basis of experiments carried out with the market simulation.

Links:
Download PDF

Monitoring of Agents for Dynamic Pricing in different Recommerce Markets

By: Nikkel Mollenhauer (@NikkelM)
Language: English

Abstract:
Sustainable recommerce markets are growing faster than ever. In such markets, customers are incentivised to resell their used products to businesses, which then refurbish and sell them again on the secondary market. However, businesses now face the challenge of having to set three different prices for the same item: One price for the new item, one for its refurbished version and the price at which items are bought back from customers. Since these prices are heavily influenced by each other, traditional pricing methods become less effective. To solve this dynamic pricing problem, a simulation framework was built which can be used to train artificial vendors to set optimised prices using Reinforcement learning algorithms. Before employing these trained agents on real markets, their fitness must be monitored and evaluated, as prices that are too high or too low can lead to high costs for the business. This thesis introduces a number of ways that such dynamic pricing agents can be monitored. We come to the conclusion that it is best to use a wide range of tools when evaluating different aspects of an agent’s performance, from running large-scale simulations to monitoring small policy changes following shifting market states.

Links:
Download PDF
Github repository

Pretraining RL-Based Pricing Agents for Recommerce Applications Using Historical market Data

By: Johann Schulze Tast (@blackjack2693)
Language: German
German Title: Vortrainieren RL-basierter Preisfindungsagenten für Recommerce mit geammelten Marktdaten

Abstract:
to be inserted when complete

Links:
to be inserted when complete