01‐09‐2024 Weekly Tag Up

Attendees

• Joe
• Chi Hui

Updates

• Need to decide if we want an RL paper or traffic control paper
• Feedback mentioned an ablation study
  • Before we change direction, we should add this to our work
• Feedback also mentioned lack of baseline
  • Very difficult to find baseline for multi-objective
  • We could instead split our work into individual components and evaluate the individual parts
• After addressing feedback, we select either to dig deeper into traffic control or apply to a different problem (e.g. satellite planning)
• In our ablation study regarding reward design:
  • We have a primary objective that we need to maximize while minimizing a constraint
  • We do apply a constraint to the dataset when normalizing but we may need to change how we penalize rewards that violate the constraint
    • This could be part of ablation study (different penalty functions)
    • We show how each one changes performance
• In ablation study regarding online lambda updater:
  • Didn't have a huge impact on algorithm performance
  • Could change how lambda is updated
    • Instead of magnitude of reward, use the improvement rate of reward
• Adding satellite planning problem would be very beneficial to paper (if we can fit it under 25% of page)
  • Maybe difficult to fit
  • Potentially good separate paper though, need to think about problem then come up with approach though

Submission Options

• Robot experiment is strongly recommended for IROS
  • March 1st
• AAAIA
  • Around June