Motivation

As a special session of the International Symposium on Artificial Intelligence and Mathematics (ISAIM 2024, Fort Lauderdale, FL, January 8–10, 2024), we are hosting the "Deep Reinforcement Learning: Bridging Theory and Practice" workshop. Deep Reinforcement Learning (DRL) is at the forefront of AI research, driving advancements in areas from robotics to finance. While there has been significant progress in reinforcement learning both from the perspective of theoretical advancements and from the perspective of empirical results, these communities have been largely disjoint. As in other fields, a bridging of both the theoretical and empirical viewpoint would be hugely beneficial to both communities and raise interesting and pertinent research questions! The goal of this workshop is to bridge this gap and foster an understanding and appreciation of the latest advancements in deep reinforcement learning (DRL) by emphasizing the connection between theoretical foundations and practical implementations, and to address the existing challenges facing the field. By bringing together researchers doing research on theoretical results, algorithm design and finally application to downstream realistic domains, we will build collaborations and insights that can be fruitful for both parties. While the potential of DRL is vast, there are inherent challenges that need to be addressed, both in theory and in practice. This workshop aims to provide participants with a comprehensive understanding of these challenges, backed by hands-on experiences and in-depth discussions.

Our workshop aims to bring together researchers studying the theory, and practice of deep RL, and RL algorithms more broadly to consider questions such as (but not limited to):

• Sample Efficiency: DRL algorithms often require vast amounts of data to train. What strategies can make techniques more data-efficient, and practical for real world problems?
• Exploration vs. Exploitation: Balancing the need to explore the environment and exploit known strategies remains a key challenge. How can we optimize this balance to yield more practical algorithms?
• Transfer Learning: How can we leverage knowledge from one domain to enhance performance in another?
• Stability and Convergence: With deep neural networks at their core, DRL algorithms can be prone to instability. What strategies can ensure stable learning and convergence?
• Data and Scaling: Scaling has proven to be extremely effective in domains such as NLP and vision. How do these insights extend to deep RL algorithms?
• Human-in-the-loop RL: RL algorithms will be deployed around people. How can these algorithms take insights from cognitive science/behavioral methods and apply them to develop provably safe/efficient DRL methods?
• Real-world Application Barriers: Deploying DRL in real-world scenarios poses its own set of challenges, from safety concerns to integration issues. Our workshop will provide insights into navigating these challenges effectively.

More broadly, we encourage researchers from any field related to reinforcement learning whether it be theory or practice to contribute their work! We look forward to a broad-ranging and fun discussion. The workshop is open to all - researchers in academia or industry, AI practitioners, graduate students, and industry professionals interested in the cutting-edge developments of Deep Reinforcement Learning and its challenges.

Speakers

Papers

TBA

Organizers

Contact
Reach out to Abhishek Gupta (abhgupta@cs.washington.edu) or Zhaoran Wang (zhaoranwang@gmail.com) for any questions.