README.md

tum-adlr-10

TODO

• Create presentation
• Create plots for presentation
• Try to find reason that causes random exploration to perform better than random sampling shooting
  • use random exploration to sample data points for the one-step predicitive accuracy evaluation instead of just randomly sampling states from the observation space and actions from the action space (temporarily hard coded fix in one_step_pred_accuracy.py specifically for spring-mass-damper system and short horizon)
  • implement reacher environment from paper, visualize the state space exploration and compare results with those from the paper
  • try RS without MPC
    • it's better? At least got similar performance. -> RS seems to favor exploring on velocity instead of position dimension (based on trajectory plot)
  • try use the older version of RS+MPC (without parellelizing batch of trajectories)
    • I checked the code and it should be correct
  • try print out variance value and total differential entropy
  • try reset input limitation
    • based on the trajectory plot, the datapoints seems to be too dispersed, which means input are too large
    • I tried set it to [-3,3], seems to be better, unlike RE, RS is trying to "reach out" to further regions
  • visualize predictive var on the whole state space
• Set up server to run random sampling shooting + MPC

Presentation Outline

• Introduction to our topic/Introduction to the problem we are trying to solve (Yufei)
  • hard to explore, use as few as possible samples to learn physical dynamic
  • Mention the difference to traditional reinforcement learning
• Presentation of active learning and random sampling shooting via flow chart (Ben)
  • Active Learning: Algorithm 1
  • RS: Algorithm 4
• Experiments we have conducted so far including plots (Yufei)
  • Environment we use (mass-spring-damper system)
  • Model performances BNN
    • Learning curves including train and test error
  • Active Learning evaluation
    • RS implementation
    • explain the plot
    • explain the results
  • Plot for exploration efficiency
    • visualization methods
      • Save weights for le every activearning iteration to plot bayesian prediction variance (Yufei)
      • Create training plots for every ative learning iteration (Ben)
• What are the next milestones? Are there any changes to the research hypothesis or problem statement from the pro-posal? (Ben)
  • Compare to other approaches from paper (soft-actor critic)
  • (Increase the model complexity)
  • instead of scaling complexity, we might want to delve deeper into BNN and study if it is reasonable to use MC-Dropout for this task and we'll try other active learning method