Sprint #13 (July 27)

[pagand]

Project clarification: there are two projects: optimization (full autonomous) and guidance for captions. Focus on the optimization task.

• In step() func: use model to predict next observation from action and state

• Better renderer (low priority) >> using tkinter

• State vs Observation

  • State: anything that is internal (Dynamic or static)
  • Observation: what we see (measured or calculated)
  • Actions: the manipulative variables (Speed and heading)
  • Disturbances: Anything out of control (weather, time, wave, season, ...)

• Inputs and outputs for model:

  • Inputs: state, observation, actions, disturbances of current time and the previous time
  • Output: dynamic state (model should not output actions, static states, and disturbances)

• Add Mode to the output of the model

• Optimization Use learn (policy) rl model for base first

Action items:

• [x] Write exactly the name for state, observation, action, disturbance features; convert state into - dynamic and static (on GitHub)
• [x] Github update
• [x] Gym env without renderer (Due 8/3)
• [ ] Trajectory transformers (Due 8/10)
• [ ] Two weeks for other works
• [ ] Draft paper- add rewards and gym env (briefly)

[pagand]

@yim-fan any update from the optimizer?

[pagand]

@yim-fan any update from the optimizer?

[yim-fan]

@pagand For the optimizer, I did some and experiments with different rewards, my Q learning optimizer now can navigate the vessel to somewhere near the destination but not on the exactly longitude and latitude of the bay, according to our last meeting, this should be fine. Starting today I am debugging on the trajectory transformer, trying to get it working and see it that could give us some better results.