Action Gap

The "action gap" refers to the discrepancy between an agent's intended actions (based on its values or objectives) and its actual actions, arising from various factors including computational limitations, imperfect knowledge, or cognitive biases. Current research focuses on bridging this gap through techniques like knowledge distillation and reinforcement learning, employing models such as generative flow networks and deep reinforcement learning architectures to improve action selection and value estimation. This research is significant for advancing artificial intelligence safety and ethical decision-making, with applications ranging from improving the efficiency of large language models to enhancing the performance of autonomous agents in complex environments.

Papers

October 21, 2024

Survival Multiarmed Bandits with Bootstrapping Methods
Peter Veroutis, Frédéric Godin
Objective Function Boosting Algorithm Risk Averse Multi Arm Bandit Cumulative Reward Action Gap

October 14, 2024

Action Gaps and Advantages in Continuous-Time Distributional Reinforcement Learning
Harley Wiltzer, Marc G. Bellemare, David Meger, Patrick Shafto, Yash Jhaveri
Reinforcement Learning Distributional Reinforcement Learning Advantage Feedback Traditional Reinforcement Learning Value Based Reinforcement Learning Action Gap

June 5, 2024

Adversarial Moment-Matching Distillation of Large Language Models
Chen Jia
Large Language Model Knowledge Distillation Instruction Following Adversarial Text Action Gap Distribution Matching Distillation

March 25, 2024

Towards a Formalisation of Value-based Actions and Consequentialist Ethics
Adam Wyner, Tomasz Zurek, DOrota Stachura-Zurek
Computational Approach Formalization Paper Action Gap

February 7, 2024

QGFN: Controllable Greediness with Action Values
Elaine Lau, Stephen Zhewen Lu, Ling Pan, Doina Precup, Emmanuel Bengio
Reinforcement Learning Generative Flow Network Greedy SLIM Energy Based Generative Greedy Sampling Action Gap

May 29, 2023

Action valuation of on- and off-ball soccer players based on multi-agent deep reinforcement learning
Hiroshi Nakahara, Kazushi Tsutsui, Kazuya Takeda, Keisuke Fujii
Reinforcement Learning Deep Reinforcement Learning Action Space Multi Agent Deep Reinforcement Learning Action Gap Ball Scoring

March 25, 2023

Multi-view knowledge distillation transformer for human action recognition
Ying-Chen Lin, Vincent S. Tseng
Human Action Recognition Hierarchical Vision Transformer Action Gap

March 20, 2022

Robust Action Gap Increasing with Clipped Advantage Learning
Zhe Zhang, Yaozhong Gan, Xiaoyang Tan
Optimal Action Slow Convergence CLIP Training Advantage Learning Action Gap

February 24, 2022

Metacognitive Agents for Ethical Decision Support: Conceptual Model and Research Roadmap
Catriona M. Kennedy
Cognitive Bias Ethical Decision Making Conceptual Model Affective Model Action Gap