Explore social reinforcement learning in this 28-minute conference talk presented by Natasha Jaques of Google AI at IPAM's Mathematics of Collective Intelligence Workshop. Delve into the abilities required for AI assistants, focusing on social learning and the process of learning from other intelligent agents in the environment. Examine generalization and transfer in reinforcement learning, as well as adversarial environment generation. Discover the PAIRED (Protagonist Antagonist Induced Regret Environment Design) approach for constraining adversaries in impossible environments. Investigate environment generation for web navigation and task success. Gain insights into social learning with multi-agent reinforcement learning and understand how agents utilize social learning techniques.
Natasha Jaques - Social Reinforcement Learning - IPAM at UCLA
Institute for Pure & Applied Mathematics (IPAM) via YouTube
Overview
Syllabus
What abilities does an Al assistant need?
Social learning -Learning from other intelligent agents in your environment
Generalization and transfer in RL
Adversarial environment generation
Is there a more elegant way to ensure the adversary does not create impossible environments?
PAIRED (Protagonist Antagonist Induced Regret Environment Design) • Constrain the adversary using the performance of a second agent in the same environment
Environment Generation for Web Navigation
Web navigation - task success
Conclusions
Outline
Social Learning with Multi-Agent RL
Agents are able to use social learning
Taught by
Institute for Pure & Applied Mathematics (IPAM)
Related Courses
-
Author Interview - ACCEL- Evolving Curricula with Regret-Based Environment Design
-
Stanford Seminar - Toward Scalable Autonomy - Aleksandra Faust
4.0 -
RLHF: How to Learn from Human Feedback with Reinforcement Learning
-
Building Adaptive Multi-Agent AI Systems: From Theory to Implementation
-
Toward Generalizable Embodied AI for Machine Autonomy
-
Multi-Agent Dynamics with Self-Play Reinforcement Learning
