Model improvements
Zima · 2025-11-14 · 6 hours
Goal: Iterate on model architecture and training recipe
What I did:
- Recorded new (cleaner?) dataset, with more consistent action completion strategy (turn clockwise ONLY while searching for cube)
- Added action history to input of model
- Implemented action chunk prediction
- changed loss function to L1 Loss
- mainly to avoid penalizing large errors as harshly, since that would discourage the model from outputting anything but the mean
What worked:
- nothing today
What failed:
- Even with larger dataset, and including action chunk prediction + action history input, the model regressed to the mean of the distribution
- tried a variety of hyperparamters, tried action history + action chunk prediction, and nothing worked
Key learning:
- 
- action distribution is highly imbalanced, could benefit from weighting loss based on action frequency. I.E. more rare actions are weighted more heavily
- so if the model predicts incorrectly on a rare action, the loss will be very high
Next session:
- Reevaluate next steps
- discretize actions?
- increase weight of rare actions?
- undersample / oversample to even out data distribution?
Time spent: 6 hours