taios / sokoban

A Sokoban game made with python and pygame

Sokoban with automated planning

This is the solution to task 6B from the fourth NI-UMI lecture at FIT CTU in 2020/21.

CTU username: saframa6

Solution

First I have created domain and task. Latter is automatically generated. Then I used Fast Downward planning system. Tested were all available solvers provided by Fast Downward. At the time of testing there we 14 solvers available. Each solver had 30 minutes limit to find and improve its solution on the given grid. However, only 6 of them were able to find a solution. Best solution was 271 steps found by seq-sat-fdss-1 in 46 seconds.

Second I have forked and modified the existing SOKOBAN game and simulated the best path in it.

Lastly, from the output of solvers, I have created graphs describing solver performance.

Performance

Figure 1: Performance comparison of Fast Downward planning systems. Each had 30 minutes to run and improve its solution. Planners were compared by their search time and plan cost, which is the number of movements to complete given SOKOBAN game. If the solver found multiple solutions, the one with the lowest plan cost is displayed.

Figure 2: Performace results of Fast Downward planning systems. nan means that the planner did not find any solution in 30 minutes. If the solver found multiple solutions, the one with the lowest plan cost is displayed.

Video of the best solution

• seq-sat-fdss-1, 271 steps, search time 46 seconds