Vortragsdetails

Task Scheduling in Multi-Task Reinforcement Learning via Task Similarity Metrics

Multi-Task Reinforcement Learning (MTRL) enhances sample efficiency by training a single agent on multiple tasks simultaneously. A key challenge in MTRL is negative transfer, where learning one task impedes progress on another, often due to a simplicity bias towards easier tasks. While current task scheduling methods mitigate this by prioritizing tasks based on difficulty, they largely overlook the underlying relationships between tasks, potentially leading to suboptimal learning curricula.

This thesis proposes a novel approach to task scheduling in MTRL that leverages task similarity metrics. By quantifying the relationships between tasks, a scheduler can make more informed decisions to promote positive knowledge transfer. The research involves designing a similarity-informed scheduling algorithm and integrating it into a baseline MTRL framework. The proposed method will be empirically evaluated on the Meta-World benchmark, comparing its sample efficiency and asymptotic performance against traditional difficulty-based and random scheduling strategies. The primary contribution is a more sophisticated scheduling strategy that enhances learning by explicitly considering inter-task dynamics.

In der Regel sind die Vorträge Teil von Lehrveranstaltungsreihen der Universität Bremen und nicht frei zugänglich. Bei Interesse wird um Rücksprache mit dem Sekretariat unter sek-ric(at)dfki.de gebeten.

© DFKI GmbH
zuletzt geändert am 31.03.2023