Claude Formanek - Dispelling the Mirage of Progress in Offline MARL through Standardise Baselines...
Автор: RL and Agents Reading Group
Загружено: 2025-01-23
Просмотров: 98
Описание:
UoE RL Reading Group | 23 January 2025
Speaker: Claude Formanek (University of Cape Town & InstaDeep)
Title: Dispelling the Mirage of Progress in Offline MARL through Standardise Baselines, Datasets, and Evaluation
Abstract: In this talk, I present two lines of work that seek to elevate the rigour and impact of offline MARL research. First, I reveal that simple, carefully implemented baselines can often surpass or match complex state-of-the-art methods, underscoring the need for consistent evaluation protocols. Second, I show how dataset generation and usage are frequently neglected, making fair comparisons difficult. To address this, I propose a standardized repository of over 80 datasets—complete with consistent formatting, an easy-to-use API, and robust analysis tools. By adopting these best practices, we can foster greater reproducibility and reliability in offline MARL.
Link(s): https://arxiv.org/abs/2406.09068
Bio: Claude is a PhD candidate at the University of Cape Town, specializing in Offline Multi-Agent Reinforcement Learning. He is also a Research Engineer at InstaDeep, focusing on Reinforcement Learning for Industrial Optimization.
Повторяем попытку...
Доступные форматы для скачивания:
Скачать видео
-
Информация по загрузке:
