Research map / Architectures and efficiency / RNNs and normalization
The Predictron: End-To-End Learning and Planning
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 the predictron’s predictions to select initial conditions | Pool simulator · Pool, 50 episodes; 256 initial conditions per position | pocketing events: 27 vs 10 | Helped |
| Add learned internal rewards and discounts (MRP model) (ablation) | Predictron · Random mazes and Pool | RMSE | Helped |
| Add a learned lambda accumulator (ablation) | Predictron · Random mazes and Pool | RMSE | Helped |
| Weight preturn losses by their usage in the lambda-preturn (ablation) | Lambda-predictron · Random mazes and Pool | RMSE | Helped |
| Use a predictron instead of a conventional feedforward or recurrent network | Random mazes and Pool prediction tasks · Random mazes and Pool | RMSE | Helped |
| Add consistency updates on unlabelled mazes (ablation) | Full (r, gamma, lambda)-predictron · Random mazes, semi-supervised learning | prediction error | Helped |
| Use one scalar lambda shared across predictions instead of prediction-specific lambda values (ablation) | Lambda-predictron · Random mazes and Pool | prediction performance | Hurt |
| Add skip connections to conventional architectures (ablation) | Feedforward and recurrent baselines · Random mazes and Pool, varying network depth | robustness to depth | Helped |
| Use a predictron rather than a deep network, even when the deep network has more parameters | Maze and Pool prediction architectures · Random mazes and Pool | RMSE | Helped |
Closest papers by meaning
- System 1.x: Learning to Balance Fast and Slow Planning with Language Models (ICLR 2025 · similarity 0.83)
- Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models (NeurIPS 2025 · similarity 0.83)
- Amortized Planning with Large-Scale Transformers: A Case Study on Chess (NeurIPS 2024 · similarity 0.83)
- Predicting Future Actions of Reinforcement Learning Agents (NeurIPS 2024 · similarity 0.83)
- DNCs Require More Planning Steps (ICML 2024 · similarity 0.83)
- Reward-Respecting Subtasks for Model-Based Reinforcement Learning (Abstract Reprint) (AAAI 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.