{
  "id": 516936,
  "title": "[Public LB20 / Private FAIL] Write-Up",
  "content": "<p>First and foremost, I would like to express my gratitude to the organizers, the Kaggle team, and everyone who participated. I love AI, mathematics, and the Kaggle community.</p>\n<h1>[Public LB20 / Private FAIL] Write-Up</h1>\n<ul>\n<li>As many participants felt, this contest was challenging. I believe many people based their approaches on the publicly available notebook that achieved LB20 first. It was interesting to see where the differences arose.</li>\n<li>I initially focused on developing a high-performance parser. The naive parsing algorithms used in many public notebooks produced many false positives. Therefore, I adopted a stringent parser that treats negative numbers, floats, invalid range integers, and non-evaluable formats as errors. By keeping everything else the same except the parser, the score jumped from <strong>LB14 to LB19</strong>.</li>\n<li>I efficiently managed time limits using time management functions and other methods.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5927283%2Fd0dcf15611e8b1ed8a078599225b3a50%2Ftime_management.png?generation=1719528296327787&amp;alt=media\"></li>\n<li>I experimented with several sampling algorithms: the sampling algorithm from the public notebook, Thompson sampling, a simple sequential selection algorithm, and a random selection algorithm. From the perspective of accuracy and stability, I adopted the simple sequential selection algorithm. Although Thompson sampling seemed promising, its accuracy was unstable due to the design of rewards and the few trials per problem.</li>\n<li>The final base approach stabilized around public LB19~20. </li>\n<li>However, this submission failed on the private leaderboard. I have discussed my thoughts on this in the following discussion.<br>\n<a href=\"https://www.kaggle.com/competitions/ai-mathematical-olympiad-prize/discussion/516864\" target=\"_blank\">https://www.kaggle.com/competitions/ai-mathematical-olympiad-prize/discussion/516864</a></li>\n<li>Another submission was made using a completely different approach. However, since the Private LB score was not satisfactory, I will omit the explanation.</li>\n</ul>",
  "comments": [],
  "comment_count": 0,
  "messages": [
    {
      "id": 2904087,
      "postDate": "2024-07-04T07:08:38.777Z",
      "content": "<p>First and foremost, I would like to express my gratitude to the organizers, the Kaggle team, and everyone who participated. I love AI, mathematics, and the Kaggle community.</p>\n<h1>[Public LB20 / Private FAIL] Write-Up</h1>\n<ul>\n<li>As many participants felt, this contest was challenging. I believe many people based their approaches on the publicly available notebook that achieved LB20 first. It was interesting to see where the differences arose.</li>\n<li>I initially focused on developing a high-performance parser. The naive parsing algorithms used in many public notebooks produced many false positives. Therefore, I adopted a stringent parser that treats negative numbers, floats, invalid range integers, and non-evaluable formats as errors. By keeping everything else the same except the parser, the score jumped from <strong>LB14 to LB19</strong>.</li>\n<li>I efficiently managed time limits using time management functions and other methods.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5927283%2Fd0dcf15611e8b1ed8a078599225b3a50%2Ftime_management.png?generation=1719528296327787&amp;alt=media\"></li>\n<li>I experimented with several sampling algorithms: the sampling algorithm from the public notebook, Thompson sampling, a simple sequential selection algorithm, and a random selection algorithm. From the perspective of accuracy and stability, I adopted the simple sequential selection algorithm. Although Thompson sampling seemed promising, its accuracy was unstable due to the design of rewards and the few trials per problem.</li>\n<li>The final base approach stabilized around public LB19~20. </li>\n<li>However, this submission failed on the private leaderboard. I have discussed my thoughts on this in the following discussion.<br>\n<a href=\"https://www.kaggle.com/competitions/ai-mathematical-olympiad-prize/discussion/516864\" target=\"_blank\">https://www.kaggle.com/competitions/ai-mathematical-olympiad-prize/discussion/516864</a></li>\n<li>Another submission was made using a completely different approach. However, since the Private LB score was not satisfactory, I will omit the explanation.</li>\n</ul>",
      "rawMarkdown": "First and foremost, I would like to express my gratitude to the organizers, the Kaggle team, and everyone who participated. I love AI, mathematics, and the Kaggle community.\n\n# [Public LB20 / Private FAIL] Write-Up\n- As many participants felt, this contest was challenging. I believe many people based their approaches on the publicly available notebook that achieved LB20 first. It was interesting to see where the differences arose.\n- I initially focused on developing a high-performance parser. The naive parsing algorithms used in many public notebooks produced many false positives. Therefore, I adopted a stringent parser that treats negative numbers, floats, invalid range integers, and non-evaluable formats as errors. By keeping everything else the same except the parser, the score jumped from **LB14 to LB19**.\n- I efficiently managed time limits using time management functions and other methods.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5927283%2Fd0dcf15611e8b1ed8a078599225b3a50%2Ftime_management.png?generation=1719528296327787&alt=media)\n- I experimented with several sampling algorithms: the sampling algorithm from the public notebook, Thompson sampling, a simple sequential selection algorithm, and a random selection algorithm. From the perspective of accuracy and stability, I adopted the simple sequential selection algorithm. Although Thompson sampling seemed promising, its accuracy was unstable due to the design of rewards and the few trials per problem.\n- The final base approach stabilized around public LB19~20. \n- However, this submission failed on the private leaderboard. I have discussed my thoughts on this in the following discussion.\nhttps://www.kaggle.com/competitions/ai-mathematical-olympiad-prize/discussion/516864\n- Another submission was made using a completely different approach. However, since the Private LB score was not satisfactory, I will omit the explanation.",
      "votes": 10
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-20T06:53:53Z",
  "competition_id": 73231,
  "competition_slug": "ai-mathematical-olympiad-prize",
  "team_id": 11873327,
  "solution_url": "/competitions/ai-mathematical-olympiad-prize/writeups/isaka-tsuyoshi-public-lb20-private-fail-write-up"
}