Research map / Reinforcement learning / Policy gradient and actor-critic
Post: Device Placement with Cross-Entropy Minimization and Proximal Policy Optimization
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 |
|---|---|---|---|
| use Post (cross-entropy minimization plus PPO) instead of policy gradient for placement search | device placement RL · ResNet-50 4 GPUs | per-step training time reduction: 2.37 -> 0.86 s (63.7%) | Helped |
| use Post instead of policy gradient | device placement RL · Inception-V3 2/4/8 GPUs | per-step training time reduction: 39.5%, 34.6%, 51.7% | Helped |
| use Post instead of policy gradient | device placement RL · NMT 2/4/8 GPUs | per-step training time reduction: 47.0%, 47.8%, 58.9% | Helped |
| use Post instead of Expert placement | NMT placement · NMT 2-layer 2 GPUs | per-step training time (s): 2.05 -> 1.22 | Helped |
| use Post instead of Single GPU placement | Inception-V3 placement · Inception-V3 4 GPUs | per-step training time (s): 1.75 -> 1.19 | Helped |
| use Post instead of Single GPU placement | ResNet and RNNLM placement · ResNet-50 2 GPUs | per-step training time (s): 0.86 -> 0.86 | No clear effect |
| add PPO updates to cross-entropy minimization (full Post vs CE only) (ablation) | Post · ResNet-50 2 GPUs learning curve | average per-step training time during learning | Helped |
| add cross-entropy minimization to PPO (full Post vs PPO only) (ablation) | Post · ResNet-50 2 GPUs learning curve | average per-step training time during learning | Helped |
| use Post placement for end-to-end training instead of policy gradient placement | Inception-V3 4 GPUs · 40000 steps training | end-to-end training time (hours): 20.5 -> 13.5 (34.1%) | Helped |
Closest papers by meaning
- Chip Placement with Diffusion Models (ICML 2025 · similarity 0.86)
- A Structure-Aware Framework for Learning Device Placements on Computation Graphs (NeurIPS 2024 · similarity 0.85)
- Integrated Hardware Architecture and Device Placement Search (ICML 2024 · similarity 0.84)
- DIME: Diffusion-Based Maximum Entropy Reinforcement Learning (ICML 2025 · similarity 0.84)
- Reinforcement Learning within Tree Search for Fast Macro Placement (ICML 2024 · similarity 0.84)
- Reinforcement Learning Policy as Macro Regulator Rather than Macro Placer (NeurIPS 2024 · 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.