I've finally had time to check out this month TPS and I read some of the disccussions posted.
\nI read about distance metrics (1,2 ) and about the shape of the field 1 .
\nGiven those ideas I thought about 3 main things that we can extract from the data.
\nGiven the timeseries data understand it could be interesting to understand how a car hitting a ball influences the velocity vector (both magnitude and direction), this information can give us some important insight for player/ball positions/velocities from which making a goal is almost impossible.
Given the Goal position, the ball position and each player position we can compute the angle formed by a player, the ball and the goal position. We consider the ball position as the vertex of the angle, so if the angle is 0 the player is between the goal and the ball so it is quite impossible to make a goal from that position, if the angle is 180 degrees we have the ball between the player and the goal, so if no player is close by the goal is easier. A similar reasoning can be done with the player's own team goal, if the angle is 0 with your own goal probably you are trying to protect the goal making harder to score to the opposing team, if the angle is 180 you won't be able to protect the goal.
Angle formed between player velocity vector and ball velocity vector. Combined with point 1 and both object positions this can give us an idea of where the ball would go if the given player hits it.
This challenge in my opinion will be quite hard but hopefully it will be a pool of interesting ideas.
\nHappy kaggling to everyone.
I've finally had time to check out this month TPS and I read some of the disccussions posted.
\nI read about distance metrics (1,2 ) and about the shape of the field 1 .
\nGiven those ideas I thought about 3 main things that we can extract from the data.
\nGiven the timeseries data understand it could be interesting to understand how a car hitting a ball influences the velocity vector (both magnitude and direction), this information can give us some important insight for player/ball positions/velocities from which making a goal is almost impossible.
Given the Goal position, the ball position and each player position we can compute the angle formed by a player, the ball and the goal position. We consider the ball position as the vertex of the angle, so if the angle is 0 the player is between the goal and the ball so it is quite impossible to make a goal from that position, if the angle is 180 degrees we have the ball between the player and the goal, so if no player is close by the goal is easier. A similar reasoning can be done with the player's own team goal, if the angle is 0 with your own goal probably you are trying to protect the goal making harder to score to the opposing team, if the angle is 180 you won't be able to protect the goal.
Angle formed between player velocity vector and ball velocity vector. Combined with point 1 and both object positions this can give us an idea of where the ball would go if the given player hits it.
This challenge in my opinion will be quite hard but hopefully it will be a pool of interesting ideas.
\nHappy kaggling to everyone.
these are very good tips!
\ni'm not participating in this competition, but i'm participating in a project and these tips gave me a good insight. 😄
\nthanks for sharing @pietromaldini1.
", "rawMarkdown": "these are very good tips!\n\ni'm not participating in this competition, but i'm participating in a project and these tips gave me a good insight. 😄\n\nthanks for sharing @pietromaldini1.", "votes": null }, { "id": "1967877", "postDate": "10/02/2022 18:30:34", "content": "I agree @pietromaldini1, this is a very tricky challenge that can be solved using various approaches. One may resort to several methods to create features using 2-d/ 3-d representations and co-ordinate geometry and distance measures. Also kinematics can be used to judge the impact elements as well. Temporal elements can also be used herewith. Finally, one may approach this using traditional batch/ online learning options too. Also, a creative participant may use GAN to augment the data as well, adding another perspective to the problem.
\nThis is in my opinion the most involved TPS challenge I have faced! Thanks to the organizers for offering such an experience!
", "rawMarkdown": "I agree @pietromaldini1, this is a very tricky challenge that can be solved using various approaches. One may resort to several methods to create features using 2-d/ 3-d representations and co-ordinate geometry and distance measures. Also kinematics can be used to judge the impact elements as well. Temporal elements can also be used herewith. Finally, one may approach this using traditional batch/ online learning options too. Also, a creative participant may use GAN to augment the data as well, adding another perspective to the problem. \n\nThis is in my opinion the most involved TPS challenge I have faced! Thanks to the organizers for offering such an experience!", "votes": null }, { "id": "1967989", "postDate": "10/02/2022 19:34:13", "content": "Hey @pietromaldini1, It will certainly be quite hard to simulate the game with the numerous variables and unknowns we have. I have tried making a crude agent here where I try to simulate the next few iterations from only the given positions and velocities we have of the ball and the players, I'm avoiding acceleration for now as that's completely player-controlled and I don't want to end up with a complex difficult-to-test rule-based agent.
\nWhile tabulating and creating new features is certainly a good way to go about it, I feel like testing using moderately-trained RL agents that behave close enough to the given train data, and then using those agents to simulate the matches in the test data should be, even if not better, atleast more exciting :)
", "rawMarkdown": "Hey @pietromaldini1, It will certainly be quite hard to simulate the game with the numerous variables and unknowns we have. I have tried making a crude agent [here](https://www.kaggle.com/code/aatiffraz/prediction-by-simulation-lets-play-rocket-league) where I try to simulate the next few iterations from only the given positions and velocities we have of the ball and the players, I'm avoiding acceleration for now as that's completely player-controlled and I don't want to end up with a complex difficult-to-test rule-based agent.\n\nWhile tabulating and creating new features is certainly a good way to go about it, I feel like testing using moderately-trained RL agents that behave close enough to the given train data, and then using those agents to simulate the matches in the test data should be, even if not better, atleast more exciting :)", "votes": null }, { "id": "1969491", "postDate": "10/03/2022 14:15:50", "content": "great ideas, please keep sharing!
", "rawMarkdown": "great ideas, please keep sharing!", "votes": null }, { "id": "1970235", "postDate": "10/04/2022 00:42:29", "content": "Great insight @pietromaldini1
", "rawMarkdown": "Great insight @pietromaldini1", "votes": null }, { "id": "1971979", "postDate": "10/04/2022 21:22:11", "content": "This is indeed the hardest TPS so far in my opinion (at least since I am participating). It has a bunch of \"strange\" things, like the fact that we have some features present in the train data set but not on the test one. Also it does not help that I am not familiar with the Rocket League.
\nFor myself I plan to start easy and take into account the position of the ball (proximity to a goal), the players/team near it and the boost of said players. ðŸ¤
", "rawMarkdown": "This is indeed the hardest TPS so far in my opinion (at least since I am participating). It has a bunch of \"strange\" things, like the fact that we have some features present in the train data set but not on the test one. Also it does not help that I am not familiar with the Rocket League.\n\nFor myself I plan to start easy and take into account the position of the ball (proximity to a goal), the players/team near it and the boost of said players. 🤠", "votes": null }, { "id": "1972168", "postDate": "10/05/2022 03:12:53", "content": "Great insight!!
", "rawMarkdown": "Great insight!!", "votes": null }, { "id": "1975644", "postDate": "10/06/2022 23:31:51", "content": "Thanks for the ideas, @pietromaldini1, at this point, I'm focusing my efforts on more simple features, but I will see if I can develop something like you describe
", "rawMarkdown": "Thanks for the ideas, @pietromaldini1, at this point, I'm focusing my efforts on more simple features, but I will see if I can develop something like you describe", "votes": null }, { "id": "1982510", "postDate": "10/11/2022 13:13:39", "content": "Thanks your great idea !!
\nHow estimate the location of the goal ?
", "rawMarkdown": "Thanks your great idea !! \n\nHow estimate the location of the goal ?", "votes": null }, { "id": "1982636", "postDate": "10/11/2022 14:29:44", "content": "From posts like this you can see that the field has a total length of around 200. The goal is centered on the X axis.
\nSo I estimated the position as (0,100) and (0,-100) because both X and y axis are 0 centered in the dataset.
", "rawMarkdown": "From posts like [this](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/358786) you can see that the field has a total length of around 200. The goal is centered on the X axis.\n \nSo I estimated the position as (0,100) and (0,-100) because both X and y axis are 0 centered in the dataset.", "votes": null }, { "id": "1989707", "postDate": "10/16/2022 06:26:00", "content": "Thank you for your kind comment !!
", "rawMarkdown": "Thank you for your kind comment !!", "votes": null }, { "id": "2003958", "postDate": "10/25/2022 23:07:45", "content": "Very clever to think of the angle with the ball as the origin! Thanks for sharing.
", "rawMarkdown": "Very clever to think of the angle with the ball as the origin! Thanks for sharing.", "votes": null }, { "id": "2629173", "postDate": "01/31/2024 17:01:41", "content": "Your proposed ideas for extracting meaningful insights from the data are intriguing! The considerations around how a car hitting a ball influences the velocity vector, the angle formed by players, the ball, and goal positions, as well as the angle between player and ball velocity vectors, are indeed valuable perspectives.
", "rawMarkdown": "Your proposed ideas for extracting meaningful insights from the data are intriguing! The considerations around how a car hitting a ball influences the velocity vector, the angle formed by players, the ball, and goal positions, as well as the angle between player and ball velocity vectors, are indeed valuable perspectives.", "votes": null } ], "comments": [ { "id": 1967459, "author_name": "jardelnascimento", "author_url": "", "post_date": "10/02/2022 14:11:41", "content": "these are very good tips!
\ni'm not participating in this competition, but i'm participating in a project and these tips gave me a good insight. 😄
\nthanks for sharing @pietromaldini1.
", "votes": null, "replies": [] }, { "id": 1967877, "author_name": "ravi20076", "author_url": "", "post_date": "10/02/2022 18:30:34", "content": "I agree @pietromaldini1, this is a very tricky challenge that can be solved using various approaches. One may resort to several methods to create features using 2-d/ 3-d representations and co-ordinate geometry and distance measures. Also kinematics can be used to judge the impact elements as well. Temporal elements can also be used herewith. Finally, one may approach this using traditional batch/ online learning options too. Also, a creative participant may use GAN to augment the data as well, adding another perspective to the problem.
\nThis is in my opinion the most involved TPS challenge I have faced! Thanks to the organizers for offering such an experience!
", "votes": null, "replies": [] }, { "id": 1967989, "author_name": "aatiffraz", "author_url": "", "post_date": "10/02/2022 19:34:13", "content": "Hey @pietromaldini1, It will certainly be quite hard to simulate the game with the numerous variables and unknowns we have. I have tried making a crude agent here where I try to simulate the next few iterations from only the given positions and velocities we have of the ball and the players, I'm avoiding acceleration for now as that's completely player-controlled and I don't want to end up with a complex difficult-to-test rule-based agent.
\nWhile tabulating and creating new features is certainly a good way to go about it, I feel like testing using moderately-trained RL agents that behave close enough to the given train data, and then using those agents to simulate the matches in the test data should be, even if not better, atleast more exciting :)
", "votes": null, "replies": [] }, { "id": 1969491, "author_name": "nehirokcu", "author_url": "", "post_date": "10/03/2022 14:15:50", "content": "great ideas, please keep sharing!
", "votes": null, "replies": [] }, { "id": 1970235, "author_name": "therealoise", "author_url": "", "post_date": "10/04/2022 00:42:29", "content": "Great insight @pietromaldini1
", "votes": null, "replies": [] }, { "id": 1971979, "author_name": "stefancomanita", "author_url": "", "post_date": "10/04/2022 21:22:11", "content": "This is indeed the hardest TPS so far in my opinion (at least since I am participating). It has a bunch of \"strange\" things, like the fact that we have some features present in the train data set but not on the test one. Also it does not help that I am not familiar with the Rocket League.
\nFor myself I plan to start easy and take into account the position of the ball (proximity to a goal), the players/team near it and the boost of said players. ðŸ¤
", "votes": null, "replies": [] }, { "id": 1972168, "author_name": "basel99", "author_url": "", "post_date": "10/05/2022 03:12:53", "content": "Great insight!!
", "votes": null, "replies": [] }, { "id": 1975644, "author_name": "cv13j0", "author_url": "", "post_date": "10/06/2022 23:31:51", "content": "Thanks for the ideas, @pietromaldini1, at this point, I'm focusing my efforts on more simple features, but I will see if I can develop something like you describe
", "votes": null, "replies": [] }, { "id": 1982510, "author_name": "shoooono", "author_url": "", "post_date": "10/11/2022 13:13:39", "content": "Thanks your great idea !!
\nHow estimate the location of the goal ?
", "votes": null, "replies": [ { "id": 1982636, "author_name": "pietromaldini1", "author_url": "", "post_date": "10/11/2022 14:29:44", "content": "From posts like this you can see that the field has a total length of around 200. The goal is centered on the X axis.
\nSo I estimated the position as (0,100) and (0,-100) because both X and y axis are 0 centered in the dataset.
", "votes": null, "replies": [] }, { "id": 1989707, "author_name": "shoooono", "author_url": "", "post_date": "10/16/2022 06:26:00", "content": "Thank you for your kind comment !!
", "votes": null, "replies": [] } ] }, { "id": 2003958, "author_name": "cristiansanabria", "author_url": "", "post_date": "10/25/2022 23:07:45", "content": "Very clever to think of the angle with the ball as the origin! Thanks for sharing.
", "votes": null, "replies": [] }, { "id": 2629173, "author_name": "sahilfatima", "author_url": "", "post_date": "01/31/2024 17:01:41", "content": "Your proposed ideas for extracting meaningful insights from the data are intriguing! The considerations around how a car hitting a ball influences the velocity vector, the angle formed by players, the ball, and goal positions, as well as the angle between player and ball velocity vectors, are indeed valuable perspectives.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1967423": "I've finally had time to check out this month TPS and I read some of the disccussions posted. \n\nI read about distance metrics ([1](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356718),[2](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356852) ) and about the shape of the field [1](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356789) .\n\nGiven those ideas I thought about 3 main things that we can extract from the data.\n1. Given the timeseries data understand it could be interesting to understand how a car hitting a ball influences the velocity vector (both magnitude and direction), this information can give us some important insight for player/ball positions/velocities from which making a goal is almost impossible.\n\n2. Given the Goal position, the ball position and each player position we can compute the angle formed by a player, the ball and the goal position. We consider the ball position as the vertex of the angle, so if the angle is 0 the player is between the goal and the ball so it is quite impossible to make a goal from that position, if the angle is 180 degrees we have the ball between the player and the goal, so if no player is close by the goal is easier. A similar reasoning can be done with the player's own team goal, if the angle is 0 with your own goal probably you are trying to protect the goal making harder to score to the opposing team, if the angle is 180 you won't be able to protect the goal.\n\n3. Angle formed between player velocity vector and ball velocity vector. Combined with point 1 and both object positions this can give us an idea of where the ball would go if the given player hits it. \n\nThis challenge in my opinion will be quite hard but hopefully it will be a pool of interesting ideas. \nHappy kaggling to everyone.", "1967459": "these are very good tips!\n\ni'm not participating in this competition, but i'm participating in a project and these tips gave me a good insight. 😄\n\nthanks for sharing @pietromaldini1.", "1967877": "I agree @pietromaldini1, this is a very tricky challenge that can be solved using various approaches. One may resort to several methods to create features using 2-d/ 3-d representations and co-ordinate geometry and distance measures. Also kinematics can be used to judge the impact elements as well. Temporal elements can also be used herewith. Finally, one may approach this using traditional batch/ online learning options too. Also, a creative participant may use GAN to augment the data as well, adding another perspective to the problem. \n\nThis is in my opinion the most involved TPS challenge I have faced! Thanks to the organizers for offering such an experience!", "1967989": "Hey @pietromaldini1, It will certainly be quite hard to simulate the game with the numerous variables and unknowns we have. I have tried making a crude agent [here](https://www.kaggle.com/code/aatiffraz/prediction-by-simulation-lets-play-rocket-league) where I try to simulate the next few iterations from only the given positions and velocities we have of the ball and the players, I'm avoiding acceleration for now as that's completely player-controlled and I don't want to end up with a complex difficult-to-test rule-based agent.\n\nWhile tabulating and creating new features is certainly a good way to go about it, I feel like testing using moderately-trained RL agents that behave close enough to the given train data, and then using those agents to simulate the matches in the test data should be, even if not better, atleast more exciting :)", "1969491": "great ideas, please keep sharing!", "1970235": "Great insight @pietromaldini1", "1971979": "This is indeed the hardest TPS so far in my opinion (at least since I am participating). It has a bunch of \"strange\" things, like the fact that we have some features present in the train data set but not on the test one. Also it does not help that I am not familiar with the Rocket League.\n\nFor myself I plan to start easy and take into account the position of the ball (proximity to a goal), the players/team near it and the boost of said players. 🤠", "1972168": "Great insight!!", "1975644": "Thanks for the ideas, @pietromaldini1, at this point, I'm focusing my efforts on more simple features, but I will see if I can develop something like you describe", "1982510": "Thanks your great idea !! \n\nHow estimate the location of the goal ?", "1982636": "From posts like [this](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/358786) you can see that the field has a total length of around 200. The goal is centered on the X axis.\n \nSo I estimated the position as (0,100) and (0,-100) because both X and y axis are 0 centered in the dataset.", "1989707": "Thank you for your kind comment !!", "2003958": "Very clever to think of the angle with the ball as the origin! Thanks for sharing.", "2629173": "Your proposed ideas for extracting meaningful insights from the data are intriguing! The considerations around how a car hitting a ball influences the velocity vector, the angle formed by players, the ball, and goal positions, as well as the angle between player and ball velocity vectors, are indeed valuable perspectives." }, "source": "meta" }