| license: apache-2.0 | |
| tags: | |
| - reinforcement-learning | |
| - real-time-rl | |
| - alphazero | |
| - mcts | |
| - jax | |
| # Finding the Time to Think in Real-Time RL — Checkpoints | |
| Pretrained base planners and gating policies for the paper | |
| **Finding the Time to Think in Real-Time RL**. | |
| A lightweight gating policy on top of a frozen AlphaZero-style MCTS planner | |
| selects a state-dependent planning budget at each decision point, across five | |
| real-time games (Pac-Man, real-time Tetris, Snake, Speed Hex, Speed Go). | |
| - 🌐 Project page: https://aneeshers.github.io/realtime-rl/ | |
| - 📄 Paper (PDF): https://aneeshers.github.io/realtime-rl/assets/finding-the-time-to-think.pdf | |
| - 💻 Code: https://github.com/Aneeshers/realtime-rl-code | |
| ## Layout | |
| ``` | |
| checkpoints/ | |
| ├── clock/{go,hex}/{base,gating} # Speed Go / Speed Hex (pgx) | |
| └── committed_action/{pacman,snake,tetris_rt}/{base,gating} # Jumanji | |
| ``` | |
| One AlphaZero base planner + one PPO gating policy per environment. See the | |
| code repo's README for the launcher scripts that consume these. | |