I created a 3D animation of a player's car position and noticed some null values during the animation. Then I remembered a core game mechanic of Rocket League when I used to play.
\nDemolition is referred to the bumps occur when a player runs into another player at supersonic speed. The victim of a demolition will explode and combust, but will respawn near their team's goal after a few seconds. If two players run into each other both at supersonic speed and their fenders collide, they will both be demolished. In higher-level gameplay, demolitions are sometimes used tactically to prevent an opponent from scoring, or to clear the way for a teammate to score.
\nYou can view the notebook with the animation here
", "messages": [ { "id": "1964906", "postDate": "10/01/2022 01:49:16", "content": "I created a 3D animation of a player's car position and noticed some null values during the animation. Then I remembered a core game mechanic of Rocket League when I used to play.
\nDemolition is referred to the bumps occur when a player runs into another player at supersonic speed. The victim of a demolition will explode and combust, but will respawn near their team's goal after a few seconds. If two players run into each other both at supersonic speed and their fenders collide, they will both be demolished. In higher-level gameplay, demolitions are sometimes used tactically to prevent an opponent from scoring, or to clear the way for a teammate to score.
\nYou can view the notebook with the animation here
", "rawMarkdown": "I created a 3D animation of a player's car position and noticed some null values during the animation. Then I remembered a core game mechanic of Rocket League when I used to play.\n\n##What is Demolition in Rocket League?\n\nDemolition is referred to the bumps occur when a player runs into another player at supersonic speed. The victim of a demolition will explode and combust, but will respawn near their team's goal after a few seconds. If two players run into each other both at supersonic speed and their fenders collide, they will both be demolished. In higher-level gameplay, demolitions are sometimes used tactically to prevent an opponent from scoring, or to clear the way for a teammate to score.\n\nYou can view the notebook with the animation [here](https://www.kaggle.com/code/mattop/rocket-league-tps-eda/notebook)", "votes": null }, { "id": "1966450", "postDate": "10/02/2022 02:45:42", "content": "Could we potentially engineer a feature from these null values that would utilize the lack of player(s) on the field? Also, I just added a 3D line plot to the notebook visualizing a segment of each player's race paths from both teams in train_0.
", "rawMarkdown": "Could we potentially engineer a feature from these null values that would utilize the lack of player(s) on the field? Also, I just added a 3D line plot to the notebook visualizing a segment of each player's race paths from both teams in train_0.", "votes": null }, { "id": "1966524", "postDate": "10/02/2022 04:31:40", "content": "I tried the obvious binary features; correlations are in the right directions albeit weak.😐
\nX = pd.DataFrame([],index=train_0.index)\nfor i in range(6):\n X[F'p{i}_isnull'] = train_0[F'p{i}_pos_x'].isnull().astype(np.int8)\n\ndisplay((X['p0_isnull']|X['p1_isnull']|X['p2_isnull']).corr(train_0['team_A_scoring_within_10sec']))\ndisplay((X['p0_isnull']|X['p1_isnull']|X['p2_isnull']).corr(train_0['team_B_scoring_within_10sec']))\ndisplay((X['p3_isnull']|X['p4_isnull']|X['p5_isnull']).corr(train_0['team_A_scoring_within_10sec']))\ndisplay((X['p3_isnull']|X['p4_isnull']|X['p5_isnull']).corr(train_0['team_B_scoring_within_10sec']))\n-0.009979106983235795\n0.024060698532368793\n0.01614721452926487\n-0.007115157504775068\nInteresting… I tried adding a small .shift() in an attempt to improve the correlation (in Rocket League you respawn in one of your team's corners after a demolition) but had no success.
X[F'p{i}_isnull'] = train_0[F'p{i}_pos_x'].shift().isnull().astype(np.int8)
Hi Matt. 😄Thanks for sharing !
\nI know the coordinate is meaningful in this contest, so do we still need to normalize the data?👀
Now this is the kind of domain knowledge that I value 😆
", "rawMarkdown": "Now this is the kind of *domain knowledge* that I value 😆", "votes": null }, { "id": "1967775", "postDate": "10/02/2022 17:11:56", "content": "That's a good question @chal1ce! I think creating a solid CV strategy and doing some testing would be helpful in finding the answer. I see you are using StandardScaler in your notebook.
", "rawMarkdown": "That's a good question @chal1ce! I think creating a solid CV strategy and doing some testing would be helpful in finding the answer. I see you are using StandardScaler in your notebook.", "votes": null }, { "id": "1967776", "postDate": "10/02/2022 17:12:18", "content": "I forgot that demolitions were even part of the game haha
", "rawMarkdown": "I forgot that demolitions were even part of the game haha", "votes": null }, { "id": "1968169", "postDate": "10/03/2022 00:13:34", "content": "Yes, because either too high or too low data will have an impact on the model.
\nAfter I use StandardScaler to make them conform to the normal distribution, the accuracy of the model will be 0.0004 to 0.0005 higher than the original.
Nevertheless the effects of these binary features on the target variables are statistically significant (very small p-values from the proportions z-test) so including them should be beneficial to the classification task.
\nfrom statsmodels.stats.proportion import proportions_ztest\n\ndf = train_0.loc[(X['p0_isnull']|X['p1_isnull']|X['p2_isnull']).astype(bool)]\nfor team in 'AB':\n prob = train_0[F'team_{team}_scoring_within_10sec'].mean()\n display(proportions_ztest(df[F'team_{team}_scoring_within_10sec'].sum(),\\\n len(df),prob,prop_var=prob))\n\ndf = train_0.loc[(X['p3_isnull']|X['p4_isnull']|X['p5_isnull']).astype(bool)]\nfor team in 'AB':\n prob = train_0[F'team_{team}_scoring_within_10sec'].mean()\n display(proportions_ztest(df[F'team_{team}_scoring_within_10sec'].sum(),\\\n len(df),prob,prop_var=prob))\n(-14.436060507821042, 3.06888721253238e-47)\n(34.806892085356104, 1.9131495080932972e-265)\n(23.35804952768753, 1.1415816957073028e-120)\n(-10.292561673261257, 7.612373796890326e-25)\nVery nice insight, those p-values are really small!
", "rawMarkdown": "Very nice insight, those p-values are really small!", "votes": null }, { "id": "1990637", "postDate": "10/16/2022 17:04:46", "content": "That's a good question @chal1ce! I think creating a solid CV strategy and doing some testing would be helpful in finding the answer. I see you are using StandardScaler in your notebook.
", "rawMarkdown": "That's a good question @chal1ce! I think creating a solid CV strategy and doing some testing would be helpful in finding the answer. I see you are using StandardScaler in your notebook.", "votes": null } ], "comments": [ { "id": 1966450, "author_name": "mattop", "author_url": "", "post_date": "10/02/2022 02:45:42", "content": "Could we potentially engineer a feature from these null values that would utilize the lack of player(s) on the field? Also, I just added a 3D line plot to the notebook visualizing a segment of each player's race paths from both teams in train_0.
", "votes": null, "replies": [ { "id": 1966524, "author_name": "siukeitin", "author_url": "", "post_date": "10/02/2022 04:31:40", "content": "I tried the obvious binary features; correlations are in the right directions albeit weak.😐
\nX = pd.DataFrame([],index=train_0.index)\nfor i in range(6):\n X[F'p{i}_isnull'] = train_0[F'p{i}_pos_x'].isnull().astype(np.int8)\n\ndisplay((X['p0_isnull']|X['p1_isnull']|X['p2_isnull']).corr(train_0['team_A_scoring_within_10sec']))\ndisplay((X['p0_isnull']|X['p1_isnull']|X['p2_isnull']).corr(train_0['team_B_scoring_within_10sec']))\ndisplay((X['p3_isnull']|X['p4_isnull']|X['p5_isnull']).corr(train_0['team_A_scoring_within_10sec']))\ndisplay((X['p3_isnull']|X['p4_isnull']|X['p5_isnull']).corr(train_0['team_B_scoring_within_10sec']))\n-0.009979106983235795\n0.024060698532368793\n0.01614721452926487\n-0.007115157504775068\nInteresting… I tried adding a small .shift() in an attempt to improve the correlation (in Rocket League you respawn in one of your team's corners after a demolition) but had no success.
X[F'p{i}_isnull'] = train_0[F'p{i}_pos_x'].shift().isnull().astype(np.int8)
Nevertheless the effects of these binary features on the target variables are statistically significant (very small p-values from the proportions z-test) so including them should be beneficial to the classification task.
\nfrom statsmodels.stats.proportion import proportions_ztest\n\ndf = train_0.loc[(X['p0_isnull']|X['p1_isnull']|X['p2_isnull']).astype(bool)]\nfor team in 'AB':\n prob = train_0[F'team_{team}_scoring_within_10sec'].mean()\n display(proportions_ztest(df[F'team_{team}_scoring_within_10sec'].sum(),\\\n len(df),prob,prop_var=prob))\n\ndf = train_0.loc[(X['p3_isnull']|X['p4_isnull']|X['p5_isnull']).astype(bool)]\nfor team in 'AB':\n prob = train_0[F'team_{team}_scoring_within_10sec'].mean()\n display(proportions_ztest(df[F'team_{team}_scoring_within_10sec'].sum(),\\\n len(df),prob,prop_var=prob))\n(-14.436060507821042, 3.06888721253238e-47)\n(34.806892085356104, 1.9131495080932972e-265)\n(23.35804952768753, 1.1415816957073028e-120)\n(-10.292561673261257, 7.612373796890326e-25)\nVery nice insight, those p-values are really small!
", "votes": null, "replies": [] } ] }, { "id": 1967478, "author_name": "", "author_url": "", "post_date": "10/02/2022 14:21:59", "content": "Hi Matt. 😄Thanks for sharing !
\nI know the coordinate is meaningful in this contest, so do we still need to normalize the data?👀
That's a good question @chal1ce! I think creating a solid CV strategy and doing some testing would be helpful in finding the answer. I see you are using StandardScaler in your notebook.
", "votes": null, "replies": [] }, { "id": 1968169, "author_name": "", "author_url": "", "post_date": "10/03/2022 00:13:34", "content": "Yes, because either too high or too low data will have an impact on the model.
\nAfter I use StandardScaler to make them conform to the normal distribution, the accuracy of the model will be 0.0004 to 0.0005 higher than the original.
Now this is the kind of domain knowledge that I value 😆
", "votes": null, "replies": [ { "id": 1967776, "author_name": "mattop", "author_url": "", "post_date": "10/02/2022 17:12:18", "content": "I forgot that demolitions were even part of the game haha
", "votes": null, "replies": [] } ] }, { "id": 1990637, "author_name": "naveedakram1", "author_url": "", "post_date": "10/16/2022 17:04:46", "content": "That's a good question @chal1ce! I think creating a solid CV strategy and doing some testing would be helpful in finding the answer. I see you are using StandardScaler in your notebook.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1964906": "I created a 3D animation of a player's car position and noticed some null values during the animation. Then I remembered a core game mechanic of Rocket League when I used to play.\n\n##What is Demolition in Rocket League?\n\nDemolition is referred to the bumps occur when a player runs into another player at supersonic speed. The victim of a demolition will explode and combust, but will respawn near their team's goal after a few seconds. If two players run into each other both at supersonic speed and their fenders collide, they will both be demolished. In higher-level gameplay, demolitions are sometimes used tactically to prevent an opponent from scoring, or to clear the way for a teammate to score.\n\nYou can view the notebook with the animation [here](https://www.kaggle.com/code/mattop/rocket-league-tps-eda/notebook)", "1966450": "Could we potentially engineer a feature from these null values that would utilize the lack of player(s) on the field? Also, I just added a 3D line plot to the notebook visualizing a segment of each player's race paths from both teams in train_0.", "1966524": "I tried the obvious binary features; correlations are in the right directions albeit weak.😐\n```\nX = pd.DataFrame([],index=train_0.index)\nfor i in range(6):\n X[F'p{i}_isnull'] = train_0[F'p{i}_pos_x'].isnull().astype(np.int8)\n\ndisplay((X['p0_isnull']|X['p1_isnull']|X['p2_isnull']).corr(train_0['team_A_scoring_within_10sec']))\ndisplay((X['p0_isnull']|X['p1_isnull']|X['p2_isnull']).corr(train_0['team_B_scoring_within_10sec']))\ndisplay((X['p3_isnull']|X['p4_isnull']|X['p5_isnull']).corr(train_0['team_A_scoring_within_10sec']))\ndisplay((X['p3_isnull']|X['p4_isnull']|X['p5_isnull']).corr(train_0['team_B_scoring_within_10sec']))\n```\n```\n-0.009979106983235795\n0.024060698532368793\n0.01614721452926487\n-0.007115157504775068\n```", "1966623": "Interesting... I tried adding a small `.shift()` in an attempt to improve the correlation (in Rocket League you respawn in one of your team's corners after a demolition) but had no success.\n\n`X[F'p{i}_isnull'] = train_0[F'p{i}_pos_x'].shift().isnull().astype(np.int8)`", "1967478": "Hi Matt. 😄Thanks for sharing !\nI know the coordinate is meaningful in this contest, so do we still need to normalize the data?👀", "1967504": "Now this is the kind of *domain knowledge* that I value 😆", "1967775": "That's a good question @chal1ce! I think creating a solid CV strategy and doing some testing would be helpful in finding the answer. I see you are using StandardScaler in your notebook.", "1967776": "I forgot that demolitions were even part of the game haha", "1968169": "Yes, because either too high or too low data will have an impact on the model. \nAfter I use StandardScaler to make them conform to the normal distribution, the accuracy of the model will be 0.0004 to 0.0005 higher than the original.", "1968242": "Nevertheless the effects of these binary features on the target variables are statistically significant (very small p-values from the proportions z-test) so including them should be beneficial to the classification task.\n```\nfrom statsmodels.stats.proportion import proportions_ztest\n\ndf = train_0.loc[(X['p0_isnull']|X['p1_isnull']|X['p2_isnull']).astype(bool)]\nfor team in 'AB':\n prob = train_0[F'team_{team}_scoring_within_10sec'].mean()\n display(proportions_ztest(df[F'team_{team}_scoring_within_10sec'].sum(),\\\n len(df),prob,prop_var=prob))\n \ndf = train_0.loc[(X['p3_isnull']|X['p4_isnull']|X['p5_isnull']).astype(bool)]\nfor team in 'AB':\n prob = train_0[F'team_{team}_scoring_within_10sec'].mean()\n display(proportions_ztest(df[F'team_{team}_scoring_within_10sec'].sum(),\\\n len(df),prob,prop_var=prob))\n```\n```\n(-14.436060507821042, 3.06888721253238e-47)\n(34.806892085356104, 1.9131495080932972e-265)\n(23.35804952768753, 1.1415816957073028e-120)\n(-10.292561673261257, 7.612373796890326e-25)\n```", "1969621": "Very nice insight, those p-values are really small!", "1990637": "That's a good question @chal1ce! I think creating a solid CV strategy and doing some testing would be helpful in finding the answer. I see you are using StandardScaler in your notebook." }, "source": "meta" }