Abstraction and Planning

1
Abstraction and Planning in Sokoban Adi Botea,
Martin Müller, Jonathan Schaeffer University of
Alberta
?

• What is Sokoban?
• An exciting and addictive puzzle
• A man in a maze has to push several stones
• to designated squares called goal squares

• Why Planning?
• Significant progress in AI planning
• has recently been achieved
• Humans not only search but also plan
• Especially challenging, as not many results
• are published about planning in games

• Why Abstraction?
• Non abstracted planning in Sokoban has a poor
  potential
• Abstraction reduces branching factor, solution
  length, and number of deadlocks
• Humans abstract the problem too

• Why Heuristic Search Is Not Enough?
• Sokoban is a hard single agent search domain
• large branching factor
• long solutions
• deadlocks
• BUT
• Junghanns Rolling Stone is a state-of-the-art
  search based solver

How is Abstraction done?

• How is Planning done?
• Use a planner (Bacchus TLPlan)
• to solve the global abstract problem
• Planning moves mean
• Move a stone from one room or tunnel
• to other room or tunnel
• Shuffle stones inside a room so that
• the man can walk between
• two entrances

• Results
• Compare Abstract Sokoban (AS) with
• Rolling Stone (RS) best available heuristic
  search based
• Sokoban solver
• Tunnel Sokoban (TS) other good planning
  approach
• Exponential reduction of the global search space