MOS testing problem: action looks random while using POUCT planner

[khuechuong]

Hi! So I just wanted to test out my understanding by re-creating the MOS problem in a different way.

• I computed robot states and object state as 2 different numpy array, because of this instead of dim, I have an array giving me min x y and max x y.
• I used POUCT with histogram belief for object search.
• My actions were forward, turn left 45, turn right 45, and declare (equivalent to find).
• Instead of having look as an action, I made agent look all the time so that agent will always receive an observation from sensor.
• Policy just gives all actions since I don't have the look into find action limitations.
• Reward is just -1 for turn, -3 forward, -1000 for wrong Declare, +1000 for right declare if agent is within a certain distance of the object.
• num_sims = 500
• The running is a little slow but that's not the main issue.
• For simplicity, I plot and save the pose of robot and the objects every timestep.

So I set epsilon to 1 just to make sure it works first. so essentially, each object belief's pose is already pretty high after a few iterations, hence the sample_belief is gives good info right away, since the probability was basically 1. The problem is it seems the action it take is still pretty random. I tried debugging the tree but it seems the largest value of action given observation of the tree is very random as well.

Here's the link full code: https://drive.google.com/file/d/1Kul8ay_ruGrTcr2hjo2_NfcwzlntAB3N/view?usp=sharing
Just do python problem.py to run the code.
Let me know if you have any questions regarding the code. Since it is very similar to the original in terms of formatting, there's a lot of code so I wasn't sure if I should put everything on here.

Here's its pose after action, it looks very random and it keeps trying to go straight even though it can't since I only have the grid at 3 by 27.

[zkytony]

What's the observation model? What's in the field of view? Can you visualize that?