Architecture of Learning from Human Demonstration – 2008

• Demonstrations Triplets Extractor:
  • Input: nothing.
  • Description:
    • An expert will play the game.
    • Each time stamp a triplet will be generated. This triplet is consisted of <TimeStamp, GameState,SetOfActions>.
    • At the end of each time stamp we check if every goal was satisfied in that time stamp or not.
    • This process ends when the game ends.
  • Output:
    • A log file that contains a set D of demonstration triplets generated within the game.
• Raw Plan Extractor:
  • Input: Demonstration triplets.
  • Description:
    • Generating goal matrix, a 2D Matrix, where one of the dimensions is the set of goals, and the other dimension is the set of demonstration triplets.
    • Extracting the raw plans for each goal using the algorithm described in the paper.
    • Generate raw cases that consists of <Snippet,Episode<Goal,GameState,Outcome>>.
  • Output:
    • Raw cases.
• Ready-To-Use Cases Generator:
  • Input:
    • Raw cases.
  • Description:
    • Generate dependency graph for each snippet according to the algorithm in the paper.
    • Remove irrelevant actions.
    • Check all the plans to see if any plan P_i is a sub-plan of another plan P_j and substitute the actions in P_j that compose plan P_i with the goal of plan P_i.
  • Output: Ready-To-Use Cases.