Research map / Label-efficient and robust learning / Knowledge distillation
A Deep Hierarchical Approach to Lifelong Learning in Minecraft
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 |
|---|---|---|---|
| Add a pretrained Navigation 1 DSN as a reusable skill | H-DRLN · Two-room domain | success percentage: 50% -> 76% | Helped |
| Transfer the Navigation 1 DSN without additional learning | Two-room domain · Two-room domain | success percentage: 50% -> 67.65% | Helped |
| Use a distilled multi-skill network with τ=0.1 rather than an array of separate DSNs | H-DRLN with DDQN · Complex Minecraft domain | success percentage: 91±4% -> 94±4% | Helped |
| Distill multiple DSNs into one network at τ=0.1 | Navigation DSN · Navigation | success percentage: 94.6 -> 81.5 | Hurt |
| Distill multiple DSNs into one network at τ=0.1 | Pick Up DSN · Pick Up | success percentage: 100 -> 99.6 | Hurt |
| Distill multiple DSNs into one network at τ=0.1 | Break DSN · Break | success percentage: 100 -> 78.5 | Hurt |
| Distill multiple DSNs into one network at τ=0.1 | Placement DSN · Placement | success percentage: 100 -> 78.5 | Hurt |
| Set distillation temperature to τ=1 rather than τ=0.1 (ablation) | Distilled multi-skill network · Navigation, Pick Up, Break, and Placement | success percentage: 81.5 -> 78.0; 99.6 -> 83.3; 78.5 -> 73.0; 78.5 -> 73.0 | Hurt |
| Add DDQN training to the H-DRLN using a DSN array (ablation) | H-DRLN with DSN array · Complex Minecraft domain | success percentage: 85±10% -> 91±4% | Helped |
| Use H-DRLN with skills instead of DDQN alone | Complex Minecraft domain · Complex Minecraft domain | success percentage | Helped |
| Add intrinsic rewards to DDQN (ablation) | DDQN · Complex Minecraft domain | training time | Hurt |
Closest papers by meaning
- Experience-based Knowledge Correction for Robust Planning in Minecraft (ICLR 2026 · similarity 0.85)
- Self-Composing Policies for Scalable Continual Reinforcement Learning (ICML 2024 · similarity 0.84)
- Experience Transfer for Multimodal LLM Agents in Minecraft Game (CVPR 2026 · similarity 0.84)
- Skill Machines: Temporal Logic Skill Composition in Reinforcement Learning (ICLR 2024 · similarity 0.84)
- Auto MC-Reward: Automated Dense Reward Design with Large Language Models for Minecraft (CVPR 2024 · similarity 0.84)
- AssistanceZero: Scalably Solving Assistance Games (ICML 2025 · similarity 0.83)
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.