170 likes | 323 Views
SOKOBAN : Single-Agent Search term project Artificial Intelligence 2000, Spring Shin Saim Choi Dong-jin. Overview. Challenging area in sokoban Detailed implementation Experiments Result and conclusion. Challenging area in sokoban. Deadlock unsolvable problem configuration
E N D
SOKOBAN: Single-Agent Searchterm project Artificial Intelligence2000, SpringShin SaimChoi Dong-jin
Overview • Challenging area in sokoban • Detailed implementation • Experiments • Result and conclusion
Challenging area in sokoban • Deadlock • unsolvable problem configuration • irreversible moves • Search-Space Size • average : 1018 • Lower Bound • hard to estimate # of stone push needed to goal • require moving stones through and away from the goal
Detailed implementation • Deadlock • IDA* + Transposition Table • Search space size • Deadlock Table + Macro • Lower bound • Move ordering + Minmatching
IDA* • Search tree space : Linear size cf. DFS : Exponential size • Cheap stack cf. List • Limit of solution length • Earlier iterations are small : Exponentially growing size
Transposition Table • Avoiding cycles • Avoiding duplicating work • Storing recently visited entry
Minmatching • Lower bound heuristic • Minimum # of pushing stone • Each stone decide particular goal. • Minimize sum of distances • Deadlock detection
Minmatching(Cont.) • Example1
Minmatching(Cont.) • Example2
Move Ordering • Lower bound heuristic • Finding right sequence of moves
Deadlock Table • Avoiding trivial deadlock • Need off-line computation • Storing deadlock information • Problem : Computing time and space
Macro Moves • Reducing search space • Combining several actions to super action • Too many macros • Tunnel macros • One-Way Tunnel Macros • Two- Way Tunnel Macros
Tunnel Macros • Example1 • Example2
Conclusion • SOKOBAN Solution • Old technique+New technique • SOKOBAN need various heuristics and search technique. • SOKOBAN need very large search space