- Pretrain CNN autoencoder, grayscale and RGB color
- Get data no drift running aim controller baseline (150 runs)
- Noise (0,0.5,1 multipliers of 0.1 10)
- 30 no noise
- 60 noise (0.05, 2.5)
- 60 noise (0.1, 5)
- Noise (0,0.5,1 multipliers of 0.1 10)
- Multiple trained models, overfit training by a lot -->
- try to generate more data by including noise to the aim controller
- try warm start learning rate
- Solved calculating rewards to go (1)
- Train with aim controller without drift -->
- rewards of 132, 746 steps on lighthouse (fixed vel 0.5 no drift)
- rewards of 200, 679 steps on lighthouse (fixed vel 1 no drift)
- successfully learns to drive but cannot take quick turns
- Get data with/without drift running aim controller baseline
- Drift enabled/disabled
- Noise (0,0.5,1 multipliers of 0.1 10)
- 15 no noise
- 30 noise (0.05, 2.5)
- 30 noise (0.1, 5)
- From best no drift model, enable drift and train over with/without drift data -->
- rewards of 203, 675 steps on lighthouse (fixed vel 0.5)
- rewards 289, 590 steps on lighthouse (fixed vel 1)
- beats baseline which obtains 273, 605 steps on lighthouse
- beats ia 0 which obtains 243, 636 steps on lighthouse
- successfully learns to drive and drift so it can take quick turns, but does not control acceleration
- beats the baseline from which it has learned
- Train model from scratch using CNN encoder for drift, steer and acceleration
- mean rewards of 280, 598 steps on lighthouse
- Not calculating reward to go correctly, it should be final - cumulative reward
- Train with adafactor optimizator
- Augment with combined agent