Updated Results

Progress

• Implemented complete recall for COMCTS:
• Stores all examples  (observation(t),reward(t)) together with their "future" {<action(t+1),observation(t+1),reward(t+1)>,<action(t+2),observation(t+2),reward(t+2)>,...)}
• Uses this knowledge to reconstruct the tree after a split
• Implemented a 4x3 grid world:

• Actions: up, down, left right
• Observations: number of adjacent walls corrupted by some normally distributed noise with standard deviation = 0.965
• Rewards: +1 for goal state, -1 for penalty state, movement costs -0.04

 

Tiger Problem: Horizon 2

 

 Tiger Problem: Horizon 10

Grid World: max. 22 steps