Research map / Reinforcement learning / Visual and representation RL
QMDP-Net: Deep Learning for Planning under Partial Observability
Peter Karkus, David Hsu, Wee Sun Lee
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 |
|---|---|---|---|
| Embed a Bayesian filter and QMDP planning structure rather than use QMDP alone | QMDP-net · Hallway2 | success rate (%): 37.3 -> 82.9 | Improved |
| Replace the structured filter with a generic LSTM filter (ablation) | QMDP-net · Maze D-29 | success rate (%): 98.0 -> 9.8 | Hurt |
| Untie value-iteration transition weights and remove their softmax constraint (ablation) | QMDP-net · Maze S-19 | success rate (%): 93.9 -> 98.7 | Mixed |
| Untie value-iteration transition weights and remove their softmax constraint (ablation) | QMDP-net · Intel Lab | success rate (%): 94.4 -> 20.0 | Hurt |
| Increase test-time planning iterations from K=90 to K=180 (ablation) | QMDP-net trained with K=90 · Intel Lab | success rate (%): 40.8 -> 83.4 | Improved |
| Increase test-time planning iterations from K=180 to K=450 (ablation) | QMDP-net trained with K=90 · Freiburg | success rate (%): 93.2 -> 92.0 | No clear effect |
| Increase test-time planning iterations from K=90 to K=450 (ablation) | QMDP-net trained with K=90 · Belgioioso | success rate (%): 60.0 -> 95.4 | Improved |
| Increase test-time planning iterations from K=90 to K=450 (ablation) | QMDP-net trained with K=90 · MIT | success rate (%): 86.2 -> 91.4 | Improved |
| Increase test-time planning iterations from K=180 to K=450 (ablation) | QMDP-net trained with K=90 · MIT | success rate (%): 96.0 -> 91.4 | Hurt |
| Raise the cost of staying to 20 times its original value in the planning model (ablation) | QMDP policy · Maze D-29 | success rate (%): 63.2 -> 65.0 | Improved |
| Raise stay cost to 50 times and turn-right cost to 10 times their original values (ablation) | QMDP policy · Maze D-29 | success rate (%): 63.2 -> 93.0 | Improved |
| Use fixed rather than task-varying environments | QMDP-net · Fixed grid | success rate (%): 98.6 vs 98.8 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports QMDP-Net as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Information-based Value Iteration Networks for Decision Making Under Uncertainty (ICLR 2026 · better on 16 of 17 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- IPOMDP-Net: A Deep Neural Network for Partially Observable Multi-Agent Planning Using Interactive POMDPs (AAAI 2019 · similarity 0.90)
- PO-PDDL: Learning Symbolic POMDPs from Visual Demonstrations for Robot Planning Under Uncertainty (NeurIPS 2026 · similarity 0.85)
- Provable Representation with Efficient Planning for Partially Observable Reinforcement Learning (ICML 2024 · similarity 0.85)
- Rethinking Transformers in Solving POMDPs (ICML 2024 · similarity 0.85)
- Partially Observable Hierarchical Reinforcement Learning with AI Planning (Student Abstract) (AAAI 2024 · similarity 0.85)
- Learning Logic Specifications for Policy Guidance in POMDPs: an Inductive Logic Programming Approach (AAAI 2025 · similarity 0.84)
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.
