Meeting debrief:
Elena finished the preprocessing (feature selection, outlier detection, adding count down)
setup departure & arrival schedule
remove hard code
Imputing Missing SOG by rolling avg
add season and schedule type (day or midnight) as features
select adversarial trips (around 10% of the ~3000 trips)
James:
Almost finish modeling the half cheetah
Reproduce good results for the LSTM of the states for 1000 steps (each with 3 horizon and 5 history)
considered (batch_num, future_number, state_dim) as dimonsion of predicted state
trained on halfcheetah-expert-v2 offline dataset
Action items:
@tzhao-ooc (Elena):
For the next deadline (June 21) consider running and converting the seq model to python code and review the code and produce results
@jnqian99 (James):
Finish the LSTM model by
removing a(i+1), ... to have causal feasible model
train on halfcheetah-medium-v2 offline dataset
Add a light-weighted MLP head for the reward and learn reward values as separate labels (maybe multi task learning)
For the next deadline (June 21), regenerate the same plot as in paper 2 (figure in page 4 paper2) for both Auto regressive and non-auto regressive and report R2 score for both of the that is bounded and reasonable for 10-20% of test data in D4RL
@Startrixx (Jack):
Send me the work that was done for the display and an expectation of the project
Push the code that you have to your branch in our git repo inside the simulator directory
Meeting debrief: Elena finished the preprocessing (feature selection, outlier detection, adding count down)
James: Almost finish modeling the half cheetah
Action items: @tzhao-ooc (Elena):
@jnqian99 (James):
@Startrixx (Jack):