![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1788308%2F128345e42d8f0f2d70a4936c37e5e6ed%2FYukawa_coulomb_compare.svg.png?generation=1563067057464765&alt=media\")
For those who use Giba's submission:\nGiba R + data.table + Simple Features -1.012 LB \nwith score = -1.176 as a basis or/and the csv files later for blending or whatsoever.
\n\nPlease, pay attention that Yukawa potential is f=f(exp(-r)/r) and not f=f(exp(-r)). If am not wrong.
\n\nOf course, the difference is not so high, might save computation time. But just needed to mention.
\n\nSo, pay attention to this part of codecode:
\n\n\n\n\nsystem.time(daij<-aij[,.(coulomb=sum(1/d),vander=sum(1/d/d/d),yukawa=sum(exp(-d))),by=.(molecule_name,atom_index,atj)])
\n
This can increase your score to > +0.02 or even a bit more (with correct input for blending).
\n\nBelow is the comparison plot [source: wikipedia]
\n\n
For those who use Giba's submission:\nGiba R + data.table + Simple Features -1.012 LB \nwith score = -1.176 as a basis or/and the csv files later for blending or whatsoever.
\n\nPlease, pay attention that Yukawa potential is f=f(exp(-r)/r) and not f=f(exp(-r)). If am not wrong.
\n\nOf course, the difference is not so high, might save computation time. But just needed to mention.
\n\nSo, pay attention to this part of codecode:
\n\n\n\n\nsystem.time(daij<-aij[,.(coulomb=sum(1/d),vander=sum(1/d/d/d),yukawa=sum(exp(-d))),by=.(molecule_name,atom_index,atj)])
\n
This can increase your score to > +0.02 or even a bit more (with correct input for blending).
\n\nBelow is the comparison plot [source: wikipedia]
\n\n
In fact, there is a cut-off introduced in Yukawa potential where the nearer particle's influence has been favored more than particles that are farther away, moreover, the constant depends on the charges' product of two particles. I did an atom-wise approximation here: Parallelization of Coulomb/Yukawa interaction. Simply summing up all the H columns you get an approximated Yukawa potential, the ansatz used is that the interaction range cut-off is 0.5 times the maximum distance of a particle to the particle of interest.
For the most accurate one, you still have to use Dirac's equation to solve for the bond state using the e^{-r}/r as potential (solving a nonlinear eigenvalue problem which I assume counts toward quantum mechanics method which is forbidden in this method). FWIW, I approximated the Dirac equation using the perturbation method for the first 10000 atoms in 1JHC (around 80000 training samples), and used the result as features for LGB, however, the CV score does not improve using LOFO.
", "rawMarkdown": "In fact, there is a cut-off introduced in Yukawa potential where the nearer particle's influence has been favored more than particles that are farther away, moreover, the constant depends on the charges' product of two particles. I did an atom-wise approximation here: [Parallelization of Coulomb/Yukawa interaction](https://www.kaggle.com/scaomath/parallelization-of-coulomb-yukawa-interaction). Simply summing up all the `H` columns you get an approximated Yukawa potential, the ansatz used is that the interaction range cut-off is 0.5 times the maximum distance of a particle to the particle of interest. \n\nFor the most accurate one, you still have to use Dirac's equation to solve for the bond state using the e^{-r}/r as potential (solving a nonlinear eigenvalue problem which I assume counts toward quantum mechanics method which is forbidden in this method). FWIW, I approximated the Dirac equation using the perturbation method for the first 10000 atoms in 1JHC (around 80000 training samples), and used the result as features for LGB, however, the CV score does not improve using LOFO.", "votes": null }, { "id": "575032", "postDate": "07/14/2019 23:30:53", "content": "That's really interesting why the score was not affected positively.
", "rawMarkdown": "That's really interesting why the score was not affected positively.", "votes": null }, { "id": "575056", "postDate": "07/15/2019 01:54:27", "content": "What I meant to say was quantum computation vs the simple formula. Adding Yukawa potential itself will greatly boost certain types' CV.
", "rawMarkdown": "What I meant to say was quantum computation vs the simple formula. Adding Yukawa potential itself will greatly boost certain types' CV.", "votes": null }, { "id": "575430", "postDate": "07/15/2019 11:56:57", "content": "I understand that. I followed the simplicity since not so many people here are with a background in Physics. (Mine is in Theoretical Physics, Particle Physics)
", "rawMarkdown": "I understand that. I followed the simplicity since not so many people here are with a background in Physics. (Mine is in Theoretical Physics, Particle Physics)", "votes": null }, { "id": "575438", "postDate": "07/15/2019 12:00:50", "content": "@scaomath but thanks for the idea, I will check with other forces as well))
", "rawMarkdown": "scaomath but thanks for the idea, I will check with other forces as well))", "votes": null }, { "id": "577195", "postDate": "07/16/2019 13:19:40", "content": "Anyone else tested the effect of this?
", "rawMarkdown": "Anyone else tested the effect of this?", "votes": null }, { "id": "578617", "postDate": "07/18/2019 00:23:25", "content": "i got: +0.017. maybe you should message Giba just in case he missed the discussion.
", "rawMarkdown": "i got: +0.017. maybe you should message Giba just in case he missed the discussion.", "votes": null }, { "id": "578624", "postDate": "07/18/2019 00:32:56", "content": "He upvoted this, so he knows I think. Glad to know that it helped. But as you see, to reach top10 the idea should be changed vastly and this was just corrections to minor results.
", "rawMarkdown": "He upvoted this, so he knows I think. Glad to know that it helped. But as you see, to reach top10 the idea should be changed vastly and this was just corrections to minor results.", "votes": null }, { "id": "578628", "postDate": "07/18/2019 00:42:11", "content": "Agree
", "rawMarkdown": "Agree", "votes": null } ], "comments": [ { "id": 574968, "author_name": "scaomath", "author_url": "", "post_date": "07/14/2019 19:07:18", "content": "In fact, there is a cut-off introduced in Yukawa potential where the nearer particle's influence has been favored more than particles that are farther away, moreover, the constant depends on the charges' product of two particles. I did an atom-wise approximation here: Parallelization of Coulomb/Yukawa interaction. Simply summing up all the H columns you get an approximated Yukawa potential, the ansatz used is that the interaction range cut-off is 0.5 times the maximum distance of a particle to the particle of interest.
For the most accurate one, you still have to use Dirac's equation to solve for the bond state using the e^{-r}/r as potential (solving a nonlinear eigenvalue problem which I assume counts toward quantum mechanics method which is forbidden in this method). FWIW, I approximated the Dirac equation using the perturbation method for the first 10000 atoms in 1JHC (around 80000 training samples), and used the result as features for LGB, however, the CV score does not improve using LOFO.
", "votes": null, "replies": [ { "id": 575032, "author_name": "muhakabartay", "author_url": "", "post_date": "07/14/2019 23:30:53", "content": "That's really interesting why the score was not affected positively.
", "votes": null, "replies": [] }, { "id": 575056, "author_name": "scaomath", "author_url": "", "post_date": "07/15/2019 01:54:27", "content": "What I meant to say was quantum computation vs the simple formula. Adding Yukawa potential itself will greatly boost certain types' CV.
", "votes": null, "replies": [] }, { "id": 575430, "author_name": "muhakabartay", "author_url": "", "post_date": "07/15/2019 11:56:57", "content": "I understand that. I followed the simplicity since not so many people here are with a background in Physics. (Mine is in Theoretical Physics, Particle Physics)
", "votes": null, "replies": [] }, { "id": 575438, "author_name": "muhakabartay", "author_url": "", "post_date": "07/15/2019 12:00:50", "content": "@scaomath but thanks for the idea, I will check with other forces as well))
", "votes": null, "replies": [] } ] }, { "id": 577195, "author_name": "muhakabartay", "author_url": "", "post_date": "07/16/2019 13:19:40", "content": "Anyone else tested the effect of this?
", "votes": null, "replies": [ { "id": 578617, "author_name": "stocks", "author_url": "", "post_date": "07/18/2019 00:23:25", "content": "i got: +0.017. maybe you should message Giba just in case he missed the discussion.
", "votes": null, "replies": [] }, { "id": 578624, "author_name": "muhakabartay", "author_url": "", "post_date": "07/18/2019 00:32:56", "content": "He upvoted this, so he knows I think. Glad to know that it helped. But as you see, to reach top10 the idea should be changed vastly and this was just corrections to minor results.
", "votes": null, "replies": [] }, { "id": 578628, "author_name": "stocks", "author_url": "", "post_date": "07/18/2019 00:42:11", "content": "Agree
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "574494": "For those who use [Giba](https://www.kaggle.com/titericz)'s submission:\n[Giba R + data.table + Simple Features -1.012 LB](https://www.kaggle.com/titericz/giba-r-data-table-simple-features-1-17-lb) \nwith score = -1.176 as a basis or/and the csv files later for blending or whatsoever.\n\nPlease, pay attention that **Yukawa potential** is f=f(exp(-r)/r) and not f=f(exp(-r)). If am not wrong.\n\nOf course, the difference is not so high, might save computation time. But just needed to mention.\n\nSo, pay attention to this part of codecode:\n> system.time(daij<-aij[,.(coulomb=sum(1/d),vander=sum(1/d/d/d),yukawa=sum(**exp(-d)**)),by=.(molecule_name,atom_index,atj)])\n\nThis can increase your score to > **+0.02** or even a bit more (with correct input for blending).\n\n## **PS.** **Please upvote if helpful.**\n\nBelow is the comparison plot [source: wikipedia]\n\n", "574968": "In fact, there is a cut-off introduced in Yukawa potential where the nearer particle's influence has been favored more than particles that are farther away, moreover, the constant depends on the charges' product of two particles. I did an atom-wise approximation here: [Parallelization of Coulomb/Yukawa interaction](https://www.kaggle.com/scaomath/parallelization-of-coulomb-yukawa-interaction). Simply summing up all the `H` columns you get an approximated Yukawa potential, the ansatz used is that the interaction range cut-off is 0.5 times the maximum distance of a particle to the particle of interest. \n\nFor the most accurate one, you still have to use Dirac's equation to solve for the bond state using the e^{-r}/r as potential (solving a nonlinear eigenvalue problem which I assume counts toward quantum mechanics method which is forbidden in this method). FWIW, I approximated the Dirac equation using the perturbation method for the first 10000 atoms in 1JHC (around 80000 training samples), and used the result as features for LGB, however, the CV score does not improve using LOFO.", "575032": "That's really interesting why the score was not affected positively.", "575056": "What I meant to say was quantum computation vs the simple formula. Adding Yukawa potential itself will greatly boost certain types' CV.", "575430": "I understand that. I followed the simplicity since not so many people here are with a background in Physics. (Mine is in Theoretical Physics, Particle Physics)", "575438": "scaomath but thanks for the idea, I will check with other forces as well))", "577195": "Anyone else tested the effect of this?", "578617": "i got: +0.017. maybe you should message Giba just in case he missed the discussion.", "578624": "He upvoted this, so he knows I think. Glad to know that it helped. But as you see, to reach top10 the idea should be changed vastly and this was just corrections to minor results.", "578628": "Agree" }, "source": "meta" }