Explore an in-depth author interview on ACCEL: Evolving Curricula with Regret-Based Environment Design. Delve into the innovative approach of combining adversarial adaptiveness of regret-based sampling methods with level-editing capabilities for creating curricula in reinforcement learning. Gain insights on minimax regret, level selection, domain-specific knowledge requirements, and the emergence of generalization in AI agents. Discover the potential applications, challenges, and future directions of this cutting-edge research in automatic curriculum generation for multi-capable agents.
Author Interview - ACCEL- Evolving Curricula with Regret-Based Environment Design
Overview
Syllabus
- Intro
- Start of interview
- How did you get into this field?
- What is minimax regret?
- What levels does the regret objective select?
- Positive value loss correcting my mistakes
- Why is the teacher not learned?
- How much domain-specific knowledge is needed?
- What problems is this applicable to?
- Single agent vs population of agents
- Measuring and balancing level difficulty
- How does generalization emerge?
- Diving deeper into the experimental results
- What are the unsolved challenges in the field?
- Where do we go from here?
Taught by
Yannic Kilcher
