Minesweeper Solver Marina Maznikova Artificial Intelligence II Fall 2011
Transcription
Minesweeper Solver Marina Maznikova Artificial Intelligence II Fall 2011
Minesweeper Solver Marina Maznikova Artificial Intelligence II Fall 2011 Agenda 1. 2. 3. 4. 5. 6. The Minesweeper Game Theoretical Background Solving Agents Evaluation Extension Summary 2 1. The Minesweeper Game 3 1. The Minesweeper Game 4 2. Theoretical Background NP-completeness (8-SAT) Solving approaches: Constraints Reinforcement learning 5 3. Solving Agents Random agent (RA) 6 3. Solving Agents Random agent (RA) Simple agent (SA) Only mines remain around a square => mark as mine All mines around a square are marked => open neighbors 7 3. Solving Agents Greedy agent (GA) Start with (0, 0) Search for mine and safe squares If nothing found 2/3 Calculate for each square the density of mines on the squares around it For each square, take the maximum of the densities calculated for this If no opened neighbors => ½ 2/3 Open the square with the minimum “probability” of mine 1/5 8 3. Solving Agents Constraints agent (CA) Start with (0, 0) Consider only the relevant squares Model the problem as constraints problem Mine square => 1; opened square => 0 Sum of mines around opened squares Number of mines in the game 9 3. Solving Agents Constraints agent (CA) Backtracking for 3000 results Calculate how many times each square is mine Mark mine squares and open safe squares If no such squares found, open the square with the lowest sum of mines 00100000 … 01111101 11111101 12322202 10 3. Solving Agents Constraints agent (CA) Problem: Exponential algorithm Impossible to generate always all the solutions Time 11 3. Solving Agents Constraints agent (CA) Problem: Exponential algorithm Impossible to generate always all the solutions Time “Simple” constraints agent (SCA) Start with (0, 0) Search for mine and safe squares If nothing found, use constraints 12 4. Evaluation Observed variables Overall result Average squares revealed per game Time Tests (20.000 games) Agent comparison on standard board Variation of board size Variation of mine density 13 4. Evaluation: Standard Game Overall Result 5 000 98 748 0 -6 192 -5 000 50 45 46 SA GA 56 40 -10 000 -15 000 60 Average Moves 55 30 20 -18 572 -20 000 -20 000 10 -25 000 0 RA SA GA CA SCA 14 RA Problem of too many solution possibilities CA SCA 14 4. Evaluation: Standard Game Backtracking is expensive Time (minutes) 60 50.72 50 40 30 20 13.74 7.02 10 0.09 0.44 RA SA 0 GA CA SCA 15 4. Evaluation: Board Size Overall Result Better performance on larger and smaller boards Average Moves 10 000 70 5 000 60 0 50 -5 000 40 -10 000 30 -15 000 20 -20 000 10 -25 000 0 RA SA Standard GA Small CA Large SCA RA SA Standard Exception: backtracking GA CA Small Large SCA 16 4. Evaluation: Board Size Larger boards require more time, especially when backtracking is used Time (minutes) 180 160 140 120 100 80 60 40 20 0 RA SA GA Standard Small CA SCA Large 17 4. Evaluation: Mine Density The higher the mine density, the worse the performance Greedy strategy is not good for few mines Average Moves Overall Result 15 000 90 10 000 80 5 000 70 60 0 50 -5 000 40 -10 000 30 -15 000 20 -20 000 10 -25 000 0 RA 40 mines SA GA 20 mines CA 60 mines SCA RA 40 mines SA GA 20 mines CA SCA 60 mines 18 4. Evaluation: Mine Density Time (minutes) 300 CA and SCA: More mines require more time 250 200 150 100 50 0 RA RA, SA, and GA: More mines require less time SA GA 40 mines 20 mines CA SCA 60 mines 19 5. Extension Possibility to open 5% of the squares as joker RA: At the beginning of the game SA: In case no mine/safe squares are found GA: In case of two squares with the same mine “probability” CA and SCA: On start, in case of too many solutions of the problem, and when two squares are “safest” 20 5. Extension: Evaluation Overall Result Jokers improve the performance 20 000 Average Moves 100 90 80 70 60 50 40 30 20 10 0 15 000 10 000 5 000 0 -5 000 -10 000 -15 000 -20 000 -25 000 RA SA GA CA Without joker With joker SCA RA SA GA Without joker … and make unsafe strategies inefficient CA SCA With joker 21 5. Extension: Evaluation Time increases because of the more moves Time (minutes) 100 90 80 70 60 50 40 30 20 10 0 RA SA GA Without joker CA SCA With joker 22 6. Summary Backtracking is expensive and for many solution possibilities inefficient Solution: Sets The Greedy strategy is not a good strategy The higher the mine density, the difficult the game Jokers help and make unsafe move strategies inefficient 23