Was anyone able to replicate the winning solution? It appears that the solution exploited a quirk within the simulator that somehow made it advantageous to oscillate between altitudes. Can someone from Kaggle confirm (or deny) this?
\nIf it was an issue with the physics implementation in the simulator, is the issue fixed in the current version of the simulator? I would like to know if trying to decipher the milestone solution is worth the effort.
\n", "messages": [ { "id": "32546", "postDate": "10/22/2013 20:34:06", "content": "
Was anyone able to replicate the winning solution? It appears that the solution exploited a quirk within the simulator that somehow made it advantageous to oscillate between altitudes. Can someone from Kaggle confirm (or deny) this?
\nIf it was an issue with the physics implementation in the simulator, is the issue fixed in the current version of the simulator? I would like to know if trying to decipher the milestone solution is worth the effort.
\n", "rawMarkdown": "", "votes": null }, { "id": "32571", "postDate": "10/23/2013 06:10:32", "content": "
A new oscillation penalty function has been added. Basically, you have 5 free motion reversals, after that there's an added cost.
\nBe aware that flights that at cutoff that in the real world were already descending, in the simulator are considered climbing.
", "rawMarkdown": "", "votes": null }, { "id": "32591", "postDate": "10/23/2013 16:17:08", "content": "That's right. We added this penalty so as to incentivize as smooth a ride as possible for passengers, as done in real life (at least let's hope so on our future travels).
", "rawMarkdown": "", "votes": null }, { "id": "32814", "postDate": "10/29/2013 06:20:15", "content": "Joyce,
\nI am still confused. Please bear with me. Are you saying that the Physics in the simulator is correct?
\nIn other words, is it possible to save time and/or fuel by sending the plane on a repeated ascend-descend cycle? Surely, that cannot be right?!
\nI understand how it is advantageous to move up or down to a cell with favorable wind. But it looks like the milestone winning solution does this regardless of wind conditions. Maybe the savings come from keeping the plane around an altitude that is normally disallowed (for example, 40,000 feet for an eastbound flight) using frequent reversals. As the winning entry has not been released, it is hard to tell what made it efficient.
\nDoes anything prevent Kaggle from publishing flight paths from the winning entry?
", "rawMarkdown": "", "votes": null }, { "id": "32827", "postDate": "10/29/2013 15:10:00", "content": "I would also like to see, at least, a few paths from the winning solution.
\nTBH, I found the winner's description, of how he constructed the paths, disappointingly vague. Perhaps, the feeling was that even very basic details of how he exploited this quirk of the simulator wouldn't be relevant in the future.
", "rawMarkdown": "", "votes": null }, { "id": "32847", "postDate": "10/29/2013 18:12:24", "content": "Any physical model is an approximation. I was solving a mathematical task. Initially I wrote a numerical optimization algorithm. It was very unstable and converged to a different local optimum every time. Planes were often going up and down for unclear reason in my results. Then I noticed that the ascent fuel burn can be as low as the cruse fuel burn (line 32 in FuelModel.fs). Obviously if a plane can increase its altitude at no extra cost then the following descending will be cheap. Explicit exploitation of this phenomenon immediately outperformed my numerical optimization. Yes, the planes in my milestone solution often ascend and descent through the whole travel. Can a real plane get advantage from ascending-descending cycle? I guess, we need an answer of an expert. I had a hope that the next version of the flight simulator will increase the ascent fuel burn. Unfortunately, the oscillation penalty was introduced instead.
", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 32571, "author_name": "gabrieleseppi", "author_url": "", "post_date": "10/23/2013 06:10:32", "content": "A new oscillation penalty function has been added. Basically, you have 5 free motion reversals, after that there's an added cost.
\nBe aware that flights that at cutoff that in the real world were already descending, in the simulator are considered climbing.
", "votes": null, "replies": [] }, { "id": 32591, "author_name": "joycenv", "author_url": "", "post_date": "10/23/2013 16:17:08", "content": "That's right. We added this penalty so as to incentivize as smooth a ride as possible for passengers, as done in real life (at least let's hope so on our future travels).
", "votes": null, "replies": [] }, { "id": 32814, "author_name": "anlthms", "author_url": "", "post_date": "10/29/2013 06:20:15", "content": "Joyce,
\nI am still confused. Please bear with me. Are you saying that the Physics in the simulator is correct?
\nIn other words, is it possible to save time and/or fuel by sending the plane on a repeated ascend-descend cycle? Surely, that cannot be right?!
\nI understand how it is advantageous to move up or down to a cell with favorable wind. But it looks like the milestone winning solution does this regardless of wind conditions. Maybe the savings come from keeping the plane around an altitude that is normally disallowed (for example, 40,000 feet for an eastbound flight) using frequent reversals. As the winning entry has not been released, it is hard to tell what made it efficient.
\nDoes anything prevent Kaggle from publishing flight paths from the winning entry?
", "votes": null, "replies": [] }, { "id": 32827, "author_name": "glazed", "author_url": "", "post_date": "10/29/2013 15:10:00", "content": "I would also like to see, at least, a few paths from the winning solution.
\nTBH, I found the winner's description, of how he constructed the paths, disappointingly vague. Perhaps, the feeling was that even very basic details of how he exploited this quirk of the simulator wouldn't be relevant in the future.
", "votes": null, "replies": [] }, { "id": 32847, "author_name": "id118392", "author_url": "", "post_date": "10/29/2013 18:12:24", "content": "Any physical model is an approximation. I was solving a mathematical task. Initially I wrote a numerical optimization algorithm. It was very unstable and converged to a different local optimum every time. Planes were often going up and down for unclear reason in my results. Then I noticed that the ascent fuel burn can be as low as the cruse fuel burn (line 32 in FuelModel.fs). Obviously if a plane can increase its altitude at no extra cost then the following descending will be cheap. Explicit exploitation of this phenomenon immediately outperformed my numerical optimization. Yes, the planes in my milestone solution often ascend and descent through the whole travel. Can a real plane get advantage from ascending-descending cycle? I guess, we need an answer of an expert. I had a hope that the next version of the flight simulator will increase the ascent fuel burn. Unfortunately, the oscillation penalty was introduced instead.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "32546": "", "32571": "", "32591": "", "32814": "", "32827": "", "32847": "" }, "source": "meta" }