{
  "id": 55862,
  "title": "new ideas to explore",
  "url": "/competitions/trackml-particle-identification/discussion/55862",
  "author_name": "",
  "post_date": "2018-05-02T14:21:46.979935Z",
  "votes": 4,
  "comment_count": 6,
  "views": 0,
  "content": "<p>This thread is about new ideas that you can explore:</p>\n\n<ol>\n<li><p>Graph convolution network</p></li>\n<li><p>pointNet / Point-wise Convolutional Neural Network</p></li>\n<li><p>Point pairwise metric learning/embedding</p></li>\n<li><p>Path clustering  </p></li>\n<li><p>Learning affinity field in 3d space ( like the openpose paper for parsing multiple human poses )?</p></li>\n<li><p>reinforcement Learning. Formulate like navigation in 3d maze problem?</p></li>\n<li><p>Monte Carlo Tree Search?</p></li>\n</ol>",
  "messages": [
    {
      "id": "322185",
      "postDate": "05/02/2018 14:21:46",
      "content": "<p>This thread is about new ideas that you can explore:</p>\n\n<ol>\n<li><p>Graph convolution network</p></li>\n<li><p>pointNet / Point-wise Convolutional Neural Network</p></li>\n<li><p>Point pairwise metric learning/embedding</p></li>\n<li><p>Path clustering  </p></li>\n<li><p>Learning affinity field in 3d space ( like the openpose paper for parsing multiple human poses )?</p></li>\n<li><p>reinforcement Learning. Formulate like navigation in 3d maze problem?</p></li>\n<li><p>Monte Carlo Tree Search?</p></li>\n</ol>",
      "rawMarkdown": "This thread is about new ideas that you can explore:\n\n\n1. Graph convolution network\n\n2. pointNet / Point-wise Convolutional Neural Network\n\n3. Point pairwise metric learning/embedding\n\n4. Path clustering  \n\n5.  Learning affinity field in 3d space ( like the openpose paper for parsing multiple human poses )?\n\n6. reinforcement Learning. Formulate like navigation in 3d maze problem?\n\n7. Monte Carlo Tree Search?",
      "votes": null
    },
    {
      "id": "322476",
      "postDate": "05/03/2018 03:03:47",
      "content": "<p>I am not an expert in reinforcement learning. I was wondering if rl methods are useful for searching path or joining the dot. Can anyone advise? Thanks</p>",
      "rawMarkdown": "I am not an expert in reinforcement learning. I was wondering if rl methods are useful for searching path or joining the dot. Can anyone advise? Thanks",
      "votes": null
    },
    {
      "id": "322584",
      "postDate": "05/03/2018 08:38:42",
      "content": "<p>In principle, certainly yes: RL is about optimization. It could be used in many ways. The most ambitious would be defining a general policy for looking for tracks - like the Kalman filters methods, just faster ; typically it would be a one-player game à la alpha go. Or a much more narrow use, like how to optimize the parameters of any slow method to make it faster (\"optimization on a budget\"). And probably many other possibilities as well. </p>",
      "rawMarkdown": "In principle, certainly yes: RL is about optimization. It could be used in many ways. The most ambitious would be defining a general policy for looking for tracks - like the Kalman filters methods, just faster ; typically it would be a one-player game à la alpha go. Or a much more narrow use, like how to optimize the parameters of any slow method to make it faster (\"optimization on a budget\"). And probably many other possibilities as well.",
      "votes": null
    },
    {
      "id": "326684",
      "postDate": "05/10/2018 07:15:30",
      "content": "<p>embedding means:</p>\n\n<p>{x1,x2,x3  ... } --&gt; [embedded network] --&gt; {y1, y2 y3 ...}</p>\n\n<p>if xi are points on a track, then y1, y2 y3... have same values (or close in values)</p>\n\n<p>Else the values of y1, y2 y3 ...  are different.</p>",
      "rawMarkdown": "embedding means:\n\n{x1,x2,x3  ... } --&gt; [embedded network] --&gt; {y1, y2 y3 ...}\n\nif xi are points on a track, then y1, y2 y3... have same values (or close in values)\n\nElse the values of y1, y2 y3 ...  are different.",
      "votes": null
    },
    {
      "id": "327658",
      "postDate": "05/12/2018 05:38:52",
      "content": "<p>\"The Connect-The-Dots Family of Puzzles: Design and Automatic Generation\"\n<a href=\"https://fstaals.net/publications/other/pointpuzzles2014/pointpuzzles2014_local.pdf\">https://fstaals.net/publications/other/pointpuzzles2014/pointpuzzles2014_local.pdf</a></p>",
      "rawMarkdown": "\"The Connect-The-Dots Family of Puzzles: Design and Automatic Generation\"\nhttps://fstaals.net/publications/other/pointpuzzles2014/pointpuzzles2014_local.pdf",
      "votes": null
    },
    {
      "id": "327662",
      "postDate": "05/12/2018 06:02:39",
      "content": "<p>Parallel Monte Carlo Search for Hough Transform</p>\n\n<p><a href=\"http://iopscience.iop.org/article/10.1088/1742-6596/898/7/072052/pdf\">http://iopscience.iop.org/article/10.1088/1742-6596/898/7/072052/pdf</a></p>",
      "rawMarkdown": "Parallel Monte Carlo Search for Hough Transform\n\nhttp://iopscience.iop.org/article/10.1088/1742-6596/898/7/072052/pdf",
      "votes": null
    },
    {
      "id": "329778",
      "postDate": "05/17/2018 06:47:41",
      "content": "<p>Some deep learning approaches can be found here (slides): <a href=\"https://indico.cern.ch/event/567550/contributions/2629737/attachments/1511334/2357008/acat_heptrkX.pdf\">https://indico.cern.ch/event/567550/contributions/2629737/attachments/1511334/2357008/acat_heptrkX.pdf</a></p>\n\n<p>They basically cover LSTM and CNN networks.</p>",
      "rawMarkdown": "Some deep learning approaches can be found here (slides): https://indico.cern.ch/event/567550/contributions/2629737/attachments/1511334/2357008/acat_heptrkX.pdf\n\nThey basically cover LSTM and CNN networks.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 322476,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "05/03/2018 03:03:47",
      "content": "<p>I am not an expert in reinforcement learning. I was wondering if rl methods are useful for searching path or joining the dot. Can anyone advise? Thanks</p>",
      "votes": null,
      "replies": [
        {
          "id": 322584,
          "author_name": "cecilegermain",
          "author_url": "",
          "post_date": "05/03/2018 08:38:42",
          "content": "<p>In principle, certainly yes: RL is about optimization. It could be used in many ways. The most ambitious would be defining a general policy for looking for tracks - like the Kalman filters methods, just faster ; typically it would be a one-player game à la alpha go. Or a much more narrow use, like how to optimize the parameters of any slow method to make it faster (\"optimization on a budget\"). And probably many other possibilities as well. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 326684,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "05/10/2018 07:15:30",
      "content": "<p>embedding means:</p>\n\n<p>{x1,x2,x3  ... } --&gt; [embedded network] --&gt; {y1, y2 y3 ...}</p>\n\n<p>if xi are points on a track, then y1, y2 y3... have same values (or close in values)</p>\n\n<p>Else the values of y1, y2 y3 ...  are different.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 327658,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "05/12/2018 05:38:52",
      "content": "<p>\"The Connect-The-Dots Family of Puzzles: Design and Automatic Generation\"\n<a href=\"https://fstaals.net/publications/other/pointpuzzles2014/pointpuzzles2014_local.pdf\">https://fstaals.net/publications/other/pointpuzzles2014/pointpuzzles2014_local.pdf</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 327662,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "05/12/2018 06:02:39",
      "content": "<p>Parallel Monte Carlo Search for Hough Transform</p>\n\n<p><a href=\"http://iopscience.iop.org/article/10.1088/1742-6596/898/7/072052/pdf\">http://iopscience.iop.org/article/10.1088/1742-6596/898/7/072052/pdf</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 329778,
      "author_name": "demonplus",
      "author_url": "",
      "post_date": "05/17/2018 06:47:41",
      "content": "<p>Some deep learning approaches can be found here (slides): <a href=\"https://indico.cern.ch/event/567550/contributions/2629737/attachments/1511334/2357008/acat_heptrkX.pdf\">https://indico.cern.ch/event/567550/contributions/2629737/attachments/1511334/2357008/acat_heptrkX.pdf</a></p>\n\n<p>They basically cover LSTM and CNN networks.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "322185": "This thread is about new ideas that you can explore:\n\n\n1. Graph convolution network\n\n2. pointNet / Point-wise Convolutional Neural Network\n\n3. Point pairwise metric learning/embedding\n\n4. Path clustering  \n\n5.  Learning affinity field in 3d space ( like the openpose paper for parsing multiple human poses )?\n\n6. reinforcement Learning. Formulate like navigation in 3d maze problem?\n\n7. Monte Carlo Tree Search?",
    "322476": "I am not an expert in reinforcement learning. I was wondering if rl methods are useful for searching path or joining the dot. Can anyone advise? Thanks",
    "322584": "In principle, certainly yes: RL is about optimization. It could be used in many ways. The most ambitious would be defining a general policy for looking for tracks - like the Kalman filters methods, just faster ; typically it would be a one-player game à la alpha go. Or a much more narrow use, like how to optimize the parameters of any slow method to make it faster (\"optimization on a budget\"). And probably many other possibilities as well.",
    "326684": "embedding means:\n\n{x1,x2,x3  ... } --&gt; [embedded network] --&gt; {y1, y2 y3 ...}\n\nif xi are points on a track, then y1, y2 y3... have same values (or close in values)\n\nElse the values of y1, y2 y3 ...  are different.",
    "327658": "\"The Connect-The-Dots Family of Puzzles: Design and Automatic Generation\"\nhttps://fstaals.net/publications/other/pointpuzzles2014/pointpuzzles2014_local.pdf",
    "327662": "Parallel Monte Carlo Search for Hough Transform\n\nhttp://iopscience.iop.org/article/10.1088/1742-6596/898/7/072052/pdf",
    "329778": "Some deep learning approaches can be found here (slides): https://indico.cern.ch/event/567550/contributions/2629737/attachments/1511334/2357008/acat_heptrkX.pdf\n\nThey basically cover LSTM and CNN networks."
  },
  "source": "meta"
}