{
  "id": 356997,
  "title": "Simulation Modelling for Predictions",
  "url": "/competitions/tabular-playground-series-oct-2022/discussion/356997",
  "author_name": "",
  "post_date": "2022-10-02T20:29:54.441014Z",
  "votes": 6,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hey Everyone!</p>\n<p>My idea was to </p>\n<ol>\n<li>create rule-based agents for starters</li>\n<li>test them against the training dataset to see what parameters would bring those agent behaviours closest to actual player behaviours. (TO-DO) </li>\n<li>The most similar-behaving agents could then be used to 'resume' the matches in the test dataset, and calculate probabilities as to which team has the best chances of scoring, using metrics like closest distance the ball got to the goal, how much time it spent in the each teams side of the field(this one scored very bad for some reason).</li>\n</ol>\n<p>While there is a lot left to do, here: <a href=\"https://www.kaggle.com/code/aatiffraz/prediction-by-simulation-lets-play-rocket-league\" target=\"_blank\">Prediction By Simulation: Lets play Rocket League!</a></p>\n<p>While most mechanics can be extracted from the training dataset, eventually the key will be to come up with agents that can behave like actual players, and more importantly, if an actual goal is not scored, then scaling in-game variables to probabilities. </p>\n<p>So far, seems like the closest a team could get the ball to the opponent's goal is a very good indicator if a goal would actually have been scored. Rather than static rules, this work could also be loaded onto a simple ML models that can learn this relationship between distance and probability from the training dataset. I will keep updating here as I finish the TODOs one by one.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4206802%2F766ff6aa78b0a652fd8ce0dd63024af3%2FScreenshot%20from%202022-10-02%2023-28-38.png?generation=1664743009686213&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": "1968037",
      "postDate": "10/02/2022 20:29:54",
      "content": "<p>Hey Everyone!</p>\n<p>My idea was to </p>\n<ol>\n<li>create rule-based agents for starters</li>\n<li>test them against the training dataset to see what parameters would bring those agent behaviours closest to actual player behaviours. (TO-DO) </li>\n<li>The most similar-behaving agents could then be used to 'resume' the matches in the test dataset, and calculate probabilities as to which team has the best chances of scoring, using metrics like closest distance the ball got to the goal, how much time it spent in the each teams side of the field(this one scored very bad for some reason).</li>\n</ol>\n<p>While there is a lot left to do, here: <a href=\"https://www.kaggle.com/code/aatiffraz/prediction-by-simulation-lets-play-rocket-league\" target=\"_blank\">Prediction By Simulation: Lets play Rocket League!</a></p>\n<p>While most mechanics can be extracted from the training dataset, eventually the key will be to come up with agents that can behave like actual players, and more importantly, if an actual goal is not scored, then scaling in-game variables to probabilities. </p>\n<p>So far, seems like the closest a team could get the ball to the opponent's goal is a very good indicator if a goal would actually have been scored. Rather than static rules, this work could also be loaded onto a simple ML models that can learn this relationship between distance and probability from the training dataset. I will keep updating here as I finish the TODOs one by one.</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4206802%2F766ff6aa78b0a652fd8ce0dd63024af3%2FScreenshot%20from%202022-10-02%2023-28-38.png?generation=1664743009686213&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "Hey Everyone!\n\nMy idea was to \n1. create rule-based agents for starters\n2. test them against the training dataset to see what parameters would bring those agent behaviours closest to actual player behaviours. (TO-DO) \n3. The most similar-behaving agents could then be used to 'resume' the matches in the test dataset, and calculate probabilities as to which team has the best chances of scoring, using metrics like closest distance the ball got to the goal, how much time it spent in the each teams side of the field(this one scored very bad for some reason).\n\nWhile there is a lot left to do, here: [Prediction By Simulation: Lets play Rocket League!](https://www.kaggle.com/code/aatiffraz/prediction-by-simulation-lets-play-rocket-league)\n\nWhile most mechanics can be extracted from the training dataset, eventually the key will be to come up with agents that can behave like actual players, and more importantly, if an actual goal is not scored, then scaling in-game variables to probabilities. \n\nSo far, seems like the closest a team could get the ball to the opponent's goal is a very good indicator if a goal would actually have been scored. Rather than static rules, this work could also be loaded onto a simple ML models that can learn this relationship between distance and probability from the training dataset. I will keep updating here as I finish the TODOs one by one.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4206802%2F766ff6aa78b0a652fd8ce0dd63024af3%2FScreenshot%20from%202022-10-02%2023-28-38.png?generation=1664743009686213&alt=media)",
      "votes": null
    },
    {
      "id": "1968082",
      "postDate": "10/02/2022 21:02:20",
      "content": "<p>Been there, done that 👍</p>\n<p>The image of the animated chart here shows the other side (where appear Team 2/green area).<br>\nOn the Notebook, I could only see the red one.</p>",
      "rawMarkdown": "Been there, done that 👍\n\nThe image of the animated chart here shows the other side (where appear Team 2/green area).\nOn the Notebook, I could only see the red one.",
      "votes": null
    },
    {
      "id": "1968730",
      "postDate": "10/03/2022 07:25:53",
      "content": "<p>Hey <a href=\"https://www.kaggle.com/mpwolke\" target=\"_blank\">@mpwolke</a>, I've accidentally set the dimensions too large there, hence only about half the field is visible at a time. You could scroll through to see the other part. I'll fix it soon.</p>",
      "rawMarkdown": "Hey @mpwolke, I've accidentally set the dimensions too large there, hence only about half the field is visible at a time. You could scroll through to see the other part. I'll fix it soon.",
      "votes": null
    },
    {
      "id": "1970102",
      "postDate": "10/03/2022 21:01:04",
      "content": "<p>Now, I saw both teams. 👍</p>",
      "rawMarkdown": "Now, I saw both teams. 👍",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1968082,
      "author_name": "mpwolke",
      "author_url": "",
      "post_date": "10/02/2022 21:02:20",
      "content": "<p>Been there, done that 👍</p>\n<p>The image of the animated chart here shows the other side (where appear Team 2/green area).<br>\nOn the Notebook, I could only see the red one.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1968730,
          "author_name": "aatiffraz",
          "author_url": "",
          "post_date": "10/03/2022 07:25:53",
          "content": "<p>Hey <a href=\"https://www.kaggle.com/mpwolke\" target=\"_blank\">@mpwolke</a>, I've accidentally set the dimensions too large there, hence only about half the field is visible at a time. You could scroll through to see the other part. I'll fix it soon.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1970102,
          "author_name": "mpwolke",
          "author_url": "",
          "post_date": "10/03/2022 21:01:04",
          "content": "<p>Now, I saw both teams. 👍</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1968037": "Hey Everyone!\n\nMy idea was to \n1. create rule-based agents for starters\n2. test them against the training dataset to see what parameters would bring those agent behaviours closest to actual player behaviours. (TO-DO) \n3. The most similar-behaving agents could then be used to 'resume' the matches in the test dataset, and calculate probabilities as to which team has the best chances of scoring, using metrics like closest distance the ball got to the goal, how much time it spent in the each teams side of the field(this one scored very bad for some reason).\n\nWhile there is a lot left to do, here: [Prediction By Simulation: Lets play Rocket League!](https://www.kaggle.com/code/aatiffraz/prediction-by-simulation-lets-play-rocket-league)\n\nWhile most mechanics can be extracted from the training dataset, eventually the key will be to come up with agents that can behave like actual players, and more importantly, if an actual goal is not scored, then scaling in-game variables to probabilities. \n\nSo far, seems like the closest a team could get the ball to the opponent's goal is a very good indicator if a goal would actually have been scored. Rather than static rules, this work could also be loaded onto a simple ML models that can learn this relationship between distance and probability from the training dataset. I will keep updating here as I finish the TODOs one by one.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4206802%2F766ff6aa78b0a652fd8ce0dd63024af3%2FScreenshot%20from%202022-10-02%2023-28-38.png?generation=1664743009686213&alt=media)",
    "1968082": "Been there, done that 👍\n\nThe image of the animated chart here shows the other side (where appear Team 2/green area).\nOn the Notebook, I could only see the red one.",
    "1968730": "Hey @mpwolke, I've accidentally set the dimensions too large there, hence only about half the field is visible at a time. You could scroll through to see the other part. I'll fix it soon.",
    "1970102": "Now, I saw both teams. 👍"
  },
  "source": "meta"
}