Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Intrinsic Motivation and Automatic Curricula via Asymmetric Self-Play
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train Bob with asymmetric self-play rather than target-task policy gradient alone | Long hallway · Long hallway | target-task success rate | Helped |
| Replace the learned Alice policy with a random Alice policy (ablation) | Asymmetric self-play · Long hallway | target-task success rate | Hurt |
| Add count-based exploration (ablation) | Long hallway policy gradient · Long hallway | target-task success rate | Helped |
| Add repeat self-play | Neural policy gradient · Mazebase Light key | target-task reward | Helped |
| Replace the learned Alice policy with a random Alice policy (ablation) | Mazebase self-play · Mazebase Light key | target-task reward | Hurt |
| Use reverse self-play instead of repeat self-play (ablation) | Mazebase self-play with p(Light off)=0.3 · Mazebase Light key | target-task speed-up | Mixed |
| Bias self-play initial light states against the target task (ablation) | Mazebase self-play · Mazebase Light key | target-task speed-up | Hurt |
| Add reset self-play | Reinforce · Mountain Car | target-task reward | Helped |
| Train directly on target-task samples without self-play (ablation) | Reinforce · Mountain Car | target-task reward | Hurt |
| Add self-play to Reinforce or TRPO | SwimmerGather · SwimmerGather | target-task reward | Helped |
| Add repeat self-play | Reinforce · StarCraft marine training | marine reward | Helped |
| Extend StarCraft episode length from 200 to 300 (ablation) | StarCraft self-play · StarCraft marine training | marine reward | No clear effect |
Closest papers by meaning
- Automated curriculum generation through setter-solver interactions (ICLR 2020 · similarity 0.88)
- It Takes Four to Tango: Multiagent Self Play for Automatic Curriculum Generation (ICLR 2022 · similarity 0.88)
- Unsupervised Training Sequence Design: Efficient and Generalizable Agent Training (AAAI 2024 · similarity 0.87)
- A Single Goal is All You Need: Skills and Exploration Emerge from Contrastive RL without Rewards, Demonstrations, or Subgoals (ICLR 2025 · similarity 0.87)
- Motif: Intrinsic Motivation from Artificial Intelligence Feedback (ICLR 2024 · similarity 0.86)
- Heterogeneous Adversarial Play in Interactive Environments (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.