![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1181310%2Fa206067305be2fa0399bd90cfbea5294%2Fscore.png?generation=1602592849044493&alt=media\")
Just for fun, who solved all tasks with delta=1,2,3,4,5?
\nUPDATE(15 Nov):
\nJust for fun, who solved all tasks with delta=1,2,3,4,5?
\nUPDATE(15 Nov):
\nHi, I solved all delta == 1.
\n
Hi, it is awesome! Seems your solution a smarter than my brute-force code))
", "rawMarkdown": "Hi, it is awesome! Seems your solution a smarter than my brute-force code))", "votes": null }, { "id": "1048691", "postDate": "10/13/2020 18:07:10", "content": "In the end your method seems more efficient 🙃
", "rawMarkdown": "In the end your method seems more efficient 🙃", "votes": null }, { "id": "1049166", "postDate": "10/14/2020 06:57:08", "content": "I believe any delta-1 SAT solver including my PycoSAT solver will solve any delta-1 puzzle pretty easily. Just remove any optional constraints, that only needed to prioritize further solvable solutions.
", "rawMarkdown": "I believe any delta-1 SAT solver including my [PycoSAT solver](https://www.kaggle.com/elvenmonk/pycosat-exploration) will solve any delta-1 puzzle pretty easily. Just remove any optional constraints, that only needed to prioritize further solvable solutions.", "votes": null }, { "id": "1049609", "postDate": "10/14/2020 15:16:22", "content": "Agreed. I've found that the Z3 solver (used in a simple way, with no additional constraints beyond the rules of the game) averages about 380 seconds per board for delta=1 (running on my rather old laptop), and has not so far found any it can't solve.
", "rawMarkdown": "Agreed. I've found that the Z3 solver (used in a simple way, with no additional constraints beyond the rules of the game) averages about 380 seconds per board for delta=1 (running on my rather old laptop), and has not so far found any it can't solve.", "votes": null }, { "id": "1049689", "postDate": "10/14/2020 16:58:57", "content": "@johnlucas0 Wow, that's very long. PycoSAT above solves ~2 boards per second in Kaggle environment.
", "rawMarkdown": "johnlucas0 Wow, that's very long. PycoSAT above solves ~2 boards per second in Kaggle environment.", "votes": null }, { "id": "1049722", "postDate": "10/14/2020 18:08:06", "content": "I am using genetic algorithms, when reviewing the behavior with different methods I found that each method works better under certain conditions.
\nI made a comparison https://www.kaggle.com/desareca/game-of-life-genetic-algorithm-spanish between various methods. The Z3 solver works very well for small deltas.
\nAs its performance is good in different conditions, I used the best performance in each case and achieved drops from 0.1023 to 0.06189.
\nHere I went from having approx. 1000 boards with error 0 to almost 10000 with delta = 1 (I am missing 45 boards). Although for higher deltas the genetic algorithms worked better.
\nThe big problem I have is time, genetic algorithms take a long time, around 5 minutes. The zolver Z3 runs the same or worse. I hope I have time to apply the Z3 solver, combine the results and improve a bit more.
", "rawMarkdown": "I am using genetic algorithms, when reviewing the behavior with different methods I found that each method works better under certain conditions.\n\nI made a comparison [https://www.kaggle.com/desareca/game-of-life-genetic-algorithm-spanish](https://www.kaggle.com/desareca/game-of-life-genetic-algorithm-spanish) between various methods. The Z3 solver works very well for small deltas.\n\nAs its performance is good in different conditions, I used the best performance in each case and achieved drops from 0.1023 to 0.06189.\n\nHere I went from having approx. 1000 boards with error 0 to almost 10000 with delta = 1 (I am missing 45 boards). Although for higher deltas the genetic algorithms worked better.\n\nThe big problem I have is time, genetic algorithms take a long time, around 5 minutes. The zolver Z3 runs the same or worse. I hope I have time to apply the Z3 solver, combine the results and improve a bit more.", "votes": null }, { "id": "1049753", "postDate": "10/14/2020 18:50:28", "content": "Interesting - I see you made your code public, so I will try it out and see how it runs on my laptop. Thanks.
", "rawMarkdown": "Interesting - I see you made your code public, so I will try it out and see how it runs on my laptop. Thanks.", "votes": null }, { "id": "1049826", "postDate": "10/14/2020 20:07:37", "content": "In my tests with solvers, many boards(delta=1) could solve in 0.1second, but some boards need several minutes, and at least the last 20 boards I haven't solve even in 1h each. But currently, the most important advantage I have got from most difficult boards(because simple boards have solved by almost everyone) just via optimized brute-force on C++
", "rawMarkdown": "In my tests with solvers, many boards(delta=1) could solve in 0.1second, but some boards need several minutes, and at least the last 20 boards I haven't solve even in 1h each. But currently, the most important advantage I have got from most difficult boards(because simple boards have solved by almost everyone) just via optimized brute-force on C++", "votes": null }, { "id": "1050198", "postDate": "10/15/2020 06:59:53", "content": "Vlad, I ran your 20 boards through my (very slow) z3 solver, as I was interested to see if there was something different about them. The first one (52575) took 30 minutes for some reason; the remaining 19 averaged 375 seconds each (which is average for delta=1 boards).
", "rawMarkdown": "Vlad, I ran your 20 boards through my (very slow) z3 solver, as I was interested to see if there was something different about them. The first one (52575) took 30 minutes for some reason; the remaining 19 averaged 375 seconds each (which is average for delta=1 boards).", "votes": null }, { "id": "1050305", "postDate": "10/15/2020 08:53:29", "content": "@johnlucas0, yeah via @elvenmonk kernel I have solved too, it took just 90 seconds total. I must check what is magic he has inside))
", "rawMarkdown": "johnlucas0, yeah via @elvenmonk kernel I have solved too, it took just 90 seconds total. I must check what is magic he has inside))", "votes": null }, { "id": "1078270", "postDate": "11/14/2020 15:20:53", "content": "delta=2 is also solved correctly by SAT solver. For several tasks was need more than 3 hours, for the last one - 36 hours.
\nAnyone solved at least 80% of delta=3? And how then?
Congrats @knstqq, it is awesome! Our team hasn't solved 5 tasks of delta=2 yet.
", "rawMarkdown": "Congrats @knstqq, it is awesome! Our team hasn't solved 5 tasks of delta=2 yet.", "votes": null }, { "id": "1078286", "postDate": "11/14/2020 15:43:36", "content": "But my teammates solved 9135 tasks of delta=3, a long time ago via solvers. We just have been stucking with the last 130.
", "rawMarkdown": "But my teammates solved 9135 tasks of delta=3, a long time ago via solvers. We just have been stucking with the last 130.", "votes": null }, { "id": "1078363", "postDate": "11/14/2020 17:20:52", "content": "Congrats @knstqq!
\nPipeline, that I've built is not focused on tasks of specific delta, so I still have even some delta-1 tasks unsolved.
\nBelow are some of my stats:
Total solved: 41733/50000\nDelta-1 solved: 10081/10090\nDelta-2 solved: 9959/10044\nDelta-3 solved: 8984/9965\nDelta-4 solved: 7099/9869\nDelta-5 solved: 5610/10032\nScore: 0.03909299200000024\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2586962%2F91753ef410758666565c07a846a95842%2Fresults_42.png?generation=1605374442361626&alt=media\")
\nI have just finished first iteration limiting time to solve each task to 5 minutes.
\nOn the second iteration, I give 25 minutes to every unsolved task, and I hope to solve many more soon.
@knstqq Speaking on how, a lot of Delta 3-5 task are quite sparse and have cyclic sub-areas. Preserving empty space and cycles is a key in solving deep tasks.
\nYou can check out my Kernels with small-boards analysis, where I was trying to put some theories, that can be also applied to entire 25x25 board:
\nhttps://www.kaggle.com/elvenmonk/complete-4x4-board-analysis
\nhttps://www.kaggle.com/elvenmonk/complete-5x5-board-analysis
I'm pretty sure my final submissions will close the gap on delta 1-2 tasks, the question is how much 3-5 tasks will stay unsolved.
", "rawMarkdown": "knstqq Speaking on how, a lot of Delta 3-5 task are quite sparse and have cyclic sub-areas. Preserving empty space and cycles is a key in solving deep tasks.\n\nYou can check out my Kernels with small-boards analysis, where I was trying to put some theories, that can be also applied to entire 25x25 board:\nhttps://www.kaggle.com/elvenmonk/complete-4x4-board-analysis\nhttps://www.kaggle.com/elvenmonk/complete-5x5-board-analysis\n\nI'm pretty sure my final submissions will close the gap on delta 1-2 tasks, the question is how much 3-5 tasks will stay unsolved.", "votes": null } ], "comments": [ { "id": 1048380, "author_name": "ebouteillon", "author_url": "", "post_date": "10/13/2020 12:40:20", "content": "Hi, I solved all delta == 1.
\n
Hi, it is awesome! Seems your solution a smarter than my brute-force code))
", "votes": null, "replies": [] }, { "id": 1049166, "author_name": "elvenmonk", "author_url": "", "post_date": "10/14/2020 06:57:08", "content": "I believe any delta-1 SAT solver including my PycoSAT solver will solve any delta-1 puzzle pretty easily. Just remove any optional constraints, that only needed to prioritize further solvable solutions.
", "votes": null, "replies": [] }, { "id": 1049609, "author_name": "johnlucas0", "author_url": "", "post_date": "10/14/2020 15:16:22", "content": "Agreed. I've found that the Z3 solver (used in a simple way, with no additional constraints beyond the rules of the game) averages about 380 seconds per board for delta=1 (running on my rather old laptop), and has not so far found any it can't solve.
", "votes": null, "replies": [] }, { "id": 1049689, "author_name": "elvenmonk", "author_url": "", "post_date": "10/14/2020 16:58:57", "content": "@johnlucas0 Wow, that's very long. PycoSAT above solves ~2 boards per second in Kaggle environment.
", "votes": null, "replies": [] }, { "id": 1049753, "author_name": "johnlucas0", "author_url": "", "post_date": "10/14/2020 18:50:28", "content": "Interesting - I see you made your code public, so I will try it out and see how it runs on my laptop. Thanks.
", "votes": null, "replies": [] }, { "id": 1049826, "author_name": "golubev", "author_url": "", "post_date": "10/14/2020 20:07:37", "content": "In my tests with solvers, many boards(delta=1) could solve in 0.1second, but some boards need several minutes, and at least the last 20 boards I haven't solve even in 1h each. But currently, the most important advantage I have got from most difficult boards(because simple boards have solved by almost everyone) just via optimized brute-force on C++
", "votes": null, "replies": [] }, { "id": 1050198, "author_name": "johnlucas0", "author_url": "", "post_date": "10/15/2020 06:59:53", "content": "Vlad, I ran your 20 boards through my (very slow) z3 solver, as I was interested to see if there was something different about them. The first one (52575) took 30 minutes for some reason; the remaining 19 averaged 375 seconds each (which is average for delta=1 boards).
", "votes": null, "replies": [] }, { "id": 1050305, "author_name": "golubev", "author_url": "", "post_date": "10/15/2020 08:53:29", "content": "@johnlucas0, yeah via @elvenmonk kernel I have solved too, it took just 90 seconds total. I must check what is magic he has inside))
", "votes": null, "replies": [] } ] }, { "id": 1048691, "author_name": "ebouteillon", "author_url": "", "post_date": "10/13/2020 18:07:10", "content": "In the end your method seems more efficient 🙃
", "votes": null, "replies": [] }, { "id": 1049722, "author_name": "desareca", "author_url": "", "post_date": "10/14/2020 18:08:06", "content": "I am using genetic algorithms, when reviewing the behavior with different methods I found that each method works better under certain conditions.
\nI made a comparison https://www.kaggle.com/desareca/game-of-life-genetic-algorithm-spanish between various methods. The Z3 solver works very well for small deltas.
\nAs its performance is good in different conditions, I used the best performance in each case and achieved drops from 0.1023 to 0.06189.
\nHere I went from having approx. 1000 boards with error 0 to almost 10000 with delta = 1 (I am missing 45 boards). Although for higher deltas the genetic algorithms worked better.
\nThe big problem I have is time, genetic algorithms take a long time, around 5 minutes. The zolver Z3 runs the same or worse. I hope I have time to apply the Z3 solver, combine the results and improve a bit more.
", "votes": null, "replies": [] }, { "id": 1078270, "author_name": "knstqq", "author_url": "", "post_date": "11/14/2020 15:20:53", "content": "delta=2 is also solved correctly by SAT solver. For several tasks was need more than 3 hours, for the last one - 36 hours.
\nAnyone solved at least 80% of delta=3? And how then?
Congrats @knstqq, it is awesome! Our team hasn't solved 5 tasks of delta=2 yet.
", "votes": null, "replies": [] }, { "id": 1078286, "author_name": "golubev", "author_url": "", "post_date": "11/14/2020 15:43:36", "content": "But my teammates solved 9135 tasks of delta=3, a long time ago via solvers. We just have been stucking with the last 130.
", "votes": null, "replies": [] }, { "id": 1078363, "author_name": "elvenmonk", "author_url": "", "post_date": "11/14/2020 17:20:52", "content": "Congrats @knstqq!
\nPipeline, that I've built is not focused on tasks of specific delta, so I still have even some delta-1 tasks unsolved.
\nBelow are some of my stats:
Total solved: 41733/50000\nDelta-1 solved: 10081/10090\nDelta-2 solved: 9959/10044\nDelta-3 solved: 8984/9965\nDelta-4 solved: 7099/9869\nDelta-5 solved: 5610/10032\nScore: 0.03909299200000024\n
\nI have just finished first iteration limiting time to solve each task to 5 minutes.
\nOn the second iteration, I give 25 minutes to every unsolved task, and I hope to solve many more soon.
@knstqq Speaking on how, a lot of Delta 3-5 task are quite sparse and have cyclic sub-areas. Preserving empty space and cycles is a key in solving deep tasks.
\nYou can check out my Kernels with small-boards analysis, where I was trying to put some theories, that can be also applied to entire 25x25 board:
\nhttps://www.kaggle.com/elvenmonk/complete-4x4-board-analysis
\nhttps://www.kaggle.com/elvenmonk/complete-5x5-board-analysis
I'm pretty sure my final submissions will close the gap on delta 1-2 tasks, the question is how much 3-5 tasks will stay unsolved.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1048330": "Just for fun, who solved all tasks with delta=1,2,3,4,5?\n\nUPDATE(15 Nov): \n* delta=1 **total solved** by @ebouteillon, @knstqq, 'Deeper Systems'\n* delta=2 **total solved** by @knstqq \n* delta=3 **left 130** 'Deeper Systems'\n* delta=4 **left 2225** 'Deeper Systems'\n* delta=5 **left 3909** 'Deeper Systems'", "1048380": "Hi, I solved all delta == 1. \n\n", "1048530": "Hi, it is awesome! Seems your solution a smarter than my brute-force code))", "1048691": "In the end your method seems more efficient 🙃", "1049166": "I believe any delta-1 SAT solver including my [PycoSAT solver](https://www.kaggle.com/elvenmonk/pycosat-exploration) will solve any delta-1 puzzle pretty easily. Just remove any optional constraints, that only needed to prioritize further solvable solutions.", "1049609": "Agreed. I've found that the Z3 solver (used in a simple way, with no additional constraints beyond the rules of the game) averages about 380 seconds per board for delta=1 (running on my rather old laptop), and has not so far found any it can't solve.", "1049689": "johnlucas0 Wow, that's very long. PycoSAT above solves ~2 boards per second in Kaggle environment.", "1049722": "I am using genetic algorithms, when reviewing the behavior with different methods I found that each method works better under certain conditions.\n\nI made a comparison [https://www.kaggle.com/desareca/game-of-life-genetic-algorithm-spanish](https://www.kaggle.com/desareca/game-of-life-genetic-algorithm-spanish) between various methods. The Z3 solver works very well for small deltas.\n\nAs its performance is good in different conditions, I used the best performance in each case and achieved drops from 0.1023 to 0.06189.\n\nHere I went from having approx. 1000 boards with error 0 to almost 10000 with delta = 1 (I am missing 45 boards). Although for higher deltas the genetic algorithms worked better.\n\nThe big problem I have is time, genetic algorithms take a long time, around 5 minutes. The zolver Z3 runs the same or worse. I hope I have time to apply the Z3 solver, combine the results and improve a bit more.", "1049753": "Interesting - I see you made your code public, so I will try it out and see how it runs on my laptop. Thanks.", "1049826": "In my tests with solvers, many boards(delta=1) could solve in 0.1second, but some boards need several minutes, and at least the last 20 boards I haven't solve even in 1h each. But currently, the most important advantage I have got from most difficult boards(because simple boards have solved by almost everyone) just via optimized brute-force on C++", "1050198": "Vlad, I ran your 20 boards through my (very slow) z3 solver, as I was interested to see if there was something different about them. The first one (52575) took 30 minutes for some reason; the remaining 19 averaged 375 seconds each (which is average for delta=1 boards).", "1050305": "johnlucas0, yeah via @elvenmonk kernel I have solved too, it took just 90 seconds total. I must check what is magic he has inside))", "1078270": "delta=2 is also solved correctly by SAT solver. For several tasks was need more than 3 hours, for the last one - 36 hours.\nAnyone solved at least 80% of delta=3? And how then?", "1078280": "Congrats @knstqq, it is awesome! Our team hasn't solved 5 tasks of delta=2 yet.", "1078286": "But my teammates solved 9135 tasks of delta=3, a long time ago via solvers. We just have been stucking with the last 130.", "1078363": "Congrats @knstqq!\nPipeline, that I've built is not focused on tasks of specific delta, so I still have even some delta-1 tasks unsolved.\nBelow are some of my stats:\n```\nTotal solved: 41733/50000\nDelta-1 solved: 10081/10090\nDelta-2 solved: 9959/10044\nDelta-3 solved: 8984/9965\nDelta-4 solved: 7099/9869\nDelta-5 solved: 5610/10032\nScore: 0.03909299200000024\n```\n\nI have just finished first iteration limiting time to solve each task to 5 minutes.\nOn the second iteration, I give 25 minutes to every unsolved task, and I hope to solve many more soon.", "1088875": "knstqq Speaking on how, a lot of Delta 3-5 task are quite sparse and have cyclic sub-areas. Preserving empty space and cycles is a key in solving deep tasks.\n\nYou can check out my Kernels with small-boards analysis, where I was trying to put some theories, that can be also applied to entire 25x25 board:\nhttps://www.kaggle.com/elvenmonk/complete-4x4-board-analysis\nhttps://www.kaggle.com/elvenmonk/complete-5x5-board-analysis\n\nI'm pretty sure my final submissions will close the gap on delta 1-2 tasks, the question is how much 3-5 tasks will stay unsolved." }, "source": "meta" }