{
  "id": 73076,
  "title": "Template-based Bayesian analysis teams?",
  "url": "/competitions/PLAsTiCC-2018/discussion/73076",
  "author_name": "",
  "post_date": "2018-11-29T14:20:32.655231400Z",
  "votes": 4,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi,</p>\n\n<p>I was wondering if anyone is trying to fit each source with multi-wavelength time-evolution templates.</p>\n\n<p>I have an algorithm, collaborative nested sampling, which may be interesting to explore the parameter space, and distinguish classes with Bayesian model comparison. It works similar to MCMC, but can analyze millions of datasets simultaneously. But this requires a low-dim (&lt;10) parametric model for each class.</p>\n\n<p>Not sure it will work well because of the flux-calibration issues, but it might be a unique approach.</p>\n\n<p>Cheers,\nJohannes</p>",
  "messages": [
    {
      "id": "429898",
      "postDate": "11/29/2018 14:20:32",
      "content": "<p>Hi,</p>\n\n<p>I was wondering if anyone is trying to fit each source with multi-wavelength time-evolution templates.</p>\n\n<p>I have an algorithm, collaborative nested sampling, which may be interesting to explore the parameter space, and distinguish classes with Bayesian model comparison. It works similar to MCMC, but can analyze millions of datasets simultaneously. But this requires a low-dim (&lt;10) parametric model for each class.</p>\n\n<p>Not sure it will work well because of the flux-calibration issues, but it might be a unique approach.</p>\n\n<p>Cheers,\nJohannes</p>",
      "rawMarkdown": "Hi,\n\nI was wondering if anyone is trying to fit each source with multi-wavelength time-evolution templates.\n\nI have an algorithm, collaborative nested sampling, which may be interesting to explore the parameter space, and distinguish classes with Bayesian model comparison. It works similar to MCMC, but can analyze millions of datasets simultaneously. But this requires a low-dim (&lt;10) parametric model for each class.\n\nNot sure it will work well because of the flux-calibration issues, but it might be a unique approach.\n\nCheers,\nJohannes",
      "votes": null
    },
    {
      "id": "430059",
      "postDate": "11/29/2018 18:57:14",
      "content": "<p>Hi Johannes,</p>\n\n<p>I have been experimenting with curve fitting. I looked at PyMultiNest, since that had been used in papers I looked at, but I had problems in understanding how to use it. I therefore started with scipy.optimize.curve_fit, but this is slow on the test data and I had to relax the optimization conditions too much. I would be interested in your fast methods.</p>",
      "rawMarkdown": "Hi Johannes,\n\nI have been experimenting with curve fitting. I looked at PyMultiNest, since that had been used in papers I looked at, but I had problems in understanding how to use it. I therefore started with scipy.optimize.curve_fit, but this is slow on the test data and I had to relax the optimization conditions too much. I would be interested in your fast methods.",
      "votes": null
    },
    {
      "id": "431634",
      "postDate": "12/02/2018 16:46:42",
      "content": "<p>Maybe this tutorial can help: <a href=\"https://johannesbuchner.github.io/pymultinest-tutorial/example1.html\">https://johannesbuchner.github.io/pymultinest-tutorial/example1.html</a>\nBasically,</p>\n\n<ol>\n<li>define your parameter space by stretching a cube</li>\n<li>define the likelihood of your model (I hope you have that)</li>\n<li>run multinest</li>\n<li>use the posterior samples (<code>post_equal_weights.dat</code> file) to understand the uncertainties on each parameter</li>\n<li>use the lnZ value of this model to compare to another model (Bayes factor)</li>\n</ol>\n\n<p>It's going to be slower than scipy.optimize.curve_fit for sure, but it will give you uncertainties and hopefully work well with little data.</p>\n\n<p>You can contact me here: <a href=\"http://astrost.at\">http://astrost.at</a></p>",
      "rawMarkdown": "Maybe this tutorial can help: https://johannesbuchner.github.io/pymultinest-tutorial/example1.html\nBasically,\n\n1. define your parameter space by stretching a cube\n2. define the likelihood of your model (I hope you have that)\n3. run multinest\n4. use the posterior samples (`post_equal_weights.dat` file) to understand the uncertainties on each parameter\n5. use the lnZ value of this model to compare to another model (Bayes factor)\n\nIt's going to be slower than scipy.optimize.curve_fit for sure, but it will give you uncertainties and hopefully work well with little data.\n\nYou can contact me here: http://astrost.at",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 430059,
      "author_name": "helgith",
      "author_url": "",
      "post_date": "11/29/2018 18:57:14",
      "content": "<p>Hi Johannes,</p>\n\n<p>I have been experimenting with curve fitting. I looked at PyMultiNest, since that had been used in papers I looked at, but I had problems in understanding how to use it. I therefore started with scipy.optimize.curve_fit, but this is slow on the test data and I had to relax the optimization conditions too much. I would be interested in your fast methods.</p>",
      "votes": null,
      "replies": [
        {
          "id": 431634,
          "author_name": "jbuchner",
          "author_url": "",
          "post_date": "12/02/2018 16:46:42",
          "content": "<p>Maybe this tutorial can help: <a href=\"https://johannesbuchner.github.io/pymultinest-tutorial/example1.html\">https://johannesbuchner.github.io/pymultinest-tutorial/example1.html</a>\nBasically,</p>\n\n<ol>\n<li>define your parameter space by stretching a cube</li>\n<li>define the likelihood of your model (I hope you have that)</li>\n<li>run multinest</li>\n<li>use the posterior samples (<code>post_equal_weights.dat</code> file) to understand the uncertainties on each parameter</li>\n<li>use the lnZ value of this model to compare to another model (Bayes factor)</li>\n</ol>\n\n<p>It's going to be slower than scipy.optimize.curve_fit for sure, but it will give you uncertainties and hopefully work well with little data.</p>\n\n<p>You can contact me here: <a href=\"http://astrost.at\">http://astrost.at</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "429898": "Hi,\n\nI was wondering if anyone is trying to fit each source with multi-wavelength time-evolution templates.\n\nI have an algorithm, collaborative nested sampling, which may be interesting to explore the parameter space, and distinguish classes with Bayesian model comparison. It works similar to MCMC, but can analyze millions of datasets simultaneously. But this requires a low-dim (&lt;10) parametric model for each class.\n\nNot sure it will work well because of the flux-calibration issues, but it might be a unique approach.\n\nCheers,\nJohannes",
    "430059": "Hi Johannes,\n\nI have been experimenting with curve fitting. I looked at PyMultiNest, since that had been used in papers I looked at, but I had problems in understanding how to use it. I therefore started with scipy.optimize.curve_fit, but this is slow on the test data and I had to relax the optimization conditions too much. I would be interested in your fast methods.",
    "431634": "Maybe this tutorial can help: https://johannesbuchner.github.io/pymultinest-tutorial/example1.html\nBasically,\n\n1. define your parameter space by stretching a cube\n2. define the likelihood of your model (I hope you have that)\n3. run multinest\n4. use the posterior samples (`post_equal_weights.dat` file) to understand the uncertainties on each parameter\n5. use the lnZ value of this model to compare to another model (Bayes factor)\n\nIt's going to be slower than scipy.optimize.curve_fit for sure, but it will give you uncertainties and hopefully work well with little data.\n\nYou can contact me here: http://astrost.at"
  },
  "source": "meta"
}