Research domain design project
English translation
Weeks 7–8: introduce and finalize topics · Week 12: review experiment plans · Weeks 13–15: implementation, evaluation, and presentation
Choose a design topic
- Choose your own research problem as an open topic or select from the suggested fields such as control, power, and robotics. You may use existing public environments or data.
- First describe the decision or control you want to improve and the current baseline method. Limit the environment, data, and equipment to what you can actually obtain.
Problem and comparison baseline
- Define state and observation, action, reward or evaluation objective, and constraints. When using only public data, distinguish what can be learned and evaluated from what cannot be verified.
- Choose a baseline policy, controller, or model and one proposed method, and specify metrics and failure conditions for assessing improvement.
Review the experiment plan
- Specify training/evaluation environments or data splits, comparison methods, training budget, and evaluation conditions.
- Prepare a baseline experiment that runs on CPU first. Add GPU, Jetson, or robot equipment when required by the topic.
Baseline implementation
- Run the baseline method with the available environment and data, and save the execution command, settings, seed, and results together.
- Organize the code so results can be checked again without changing the environment or data.
Comparative experiments and failure analysis
- Apply one improvement and compare it with the baseline under the same evaluation conditions. If result variability affects the conclusion, check multiple seeds or evaluation episodes.
- Alongside successful cases, identify conditions with degraded performance or constraint violations and analyze their causes.
Prepare results and present
- Organize executable code and configuration, comparison tables and key plots, and materials explaining the problem, method, results, and limitations.
- Cite the public environments, data, and existing implementations used, and distinguish your changes from verified contributions.