Research map / Reinforcement learning / Visual and representation RL
GALOIS: Boosting Deep Reinforcement Learning via Generalizable Logic Synthesis
Yushi Cao, Zhiming Li, Tianpei Yang, Hao Zhang, YAN ZHENG, Yi Li, Jianye HAO, Yang Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 a hierarchical white-box program with explicit cause-effect logic instead of neural DRL or fixed-sequence program synthesis | GALOIS · Complex MiniGrid tasks, including BoxKey, UnlockPickup, and MultiRoom | average return | Improved |
| Use explicit cause-effect program logic rather than a fixed sequence of subprograms | GALOIS · BoxKey semantic modification | average return: 0.119±0.020 -> 0.976±0.001 | Improved |
| Use explicit cause-effect program logic rather than a fixed sequence of subprograms | GALOIS · UnlockPickup semantic modification | average return: 0.000±0.000 -> 0.983±0.003 | Improved |
| Represent observations as valuation vectors rather than raw environment observations | neural DRL baselines · MiniGrid test environments with different sizes | average return | Improved |
| Add hierarchical program structure to policy learning | program synthesis and hierarchical methods · UnlockPickup and BoxKey | learning speed | Improved |
| Reuse a learned GALOIS knowledge base from a task with overlapping logic as initialization | GALOIS · DoorKey to BoxKey and other related environments | learning efficiency | Improved |
| Reuse only the WHERE subprogram from a related task | GALOIS · BoxKey transfer | learning efficiency | Improved |
| Reuse the BoxKey toggle rule in UnlockPickup, where the object needs to be picked up instead | GALOIS · BoxKey to UnlockPickup transfer | task completion | Hurt |
| Use valuation-vector inputs instead of raw observations | DQN, SAC, PPO, and hDQN · DoorKey 8*8 training | average return: 0.473±0.130 -> 0.919±0.071; 0.966±0.019 -> 0.938±0.052; 0.919±0.017 -> 0.958±0.008; 0.979±0.002 -> 0.928±0.114 | Mixed |
Closest papers by meaning
- DeepLTL: Learning to Efficiently Satisfy Complex LTL Specifications for Multi-Task RL (ICLR 2025 · similarity 0.88)
- $\mathcal{B}$-Coder: Value-Based Deep Reinforcement Learning for Program Synthesis (ICLR 2024 · similarity 0.88)
- One Subgoal at a Time: Zero-Shot Generalization to Arbitrary Linear Temporal Logic Requirements in Multi-Task Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Towards General-Purpose Model-Free Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Reinforcement Learning for Reasoning in Large Language Models with One Training Example (NeurIPS 2025 · similarity 0.87)
- Beyond Pass@ 1: Self-Play with Variational Problem Synthesis Sustains RLVR (ICLR 2026 · similarity 0.87)
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.
