So far, the main public kernels are using the host's starter notebook with a tiny modification. But, intrinsically, a system of AV on the roads among other agents can be seen as a distribution of points for which we're required to predict next destinations. Hence, even a classical machine learning approach could work.
\nI'm a little bit surprised that none of available approaches goes in that direction 🤔 !?
\nNevertheless, we're just at the beginning of this competition and there still, likely, lot of ideas to be discovered.
\nAnd you, what's your experimental approach for now ?
\n\nThe datasets used to train and predict with LGBM models are :
\n\n[Updates 2]
\nSo far, the main public kernels are using the host's starter notebook with a tiny modification. But, intrinsically, a system of AV on the roads among other agents can be seen as a distribution of points for which we're required to predict next destinations. Hence, even a classical machine learning approach could work.
\nI'm a little bit surprised that none of available approaches goes in that direction 🤔 !?
\nNevertheless, we're just at the beginning of this competition and there still, likely, lot of ideas to be discovered.
\nAnd you, what's your experimental approach for now ?
\n\nThe datasets used to train and predict with LGBM models are :
\n\n[Updates 2]
\nHi,
\nThanks @kneroma for the kernel. There's a similar thread where ideas like constant velocity Kalman filter models are discussed. The problem with using only the history positions is that they don't take into account the free space (other than those occupied by surrounding vehicles) and traffic light situation. That said, these methods can be used to smooth the output (from DL models) or can be used as one candidate in multiple predictions.
", "rawMarkdown": "Hi,\n\nThanks @kneroma for the kernel. There's a [similar thread](https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/188330) where ideas like constant velocity Kalman filter models are discussed. The problem with using only the history positions is that they don't take into account the free space (other than those occupied by surrounding vehicles) and traffic light situation. That said, these methods can be used to smooth the output (from DL models) or can be used as one candidate in multiple predictions.", "votes": null } ], "comments": [ { "id": 1053639, "author_name": "suryajrrafl", "author_url": "", "post_date": "10/19/2020 07:31:13", "content": "Hi,
\nThanks @kneroma for the kernel. There's a similar thread where ideas like constant velocity Kalman filter models are discussed. The problem with using only the history positions is that they don't take into account the free space (other than those occupied by surrounding vehicles) and traffic light situation. That said, these methods can be used to smooth the output (from DL models) or can be used as one candidate in multiple predictions.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "989288": "So far, the main public kernels are using the **host's starter notebook** with a tiny modification. But, intrinsically, a system of AV on the roads among other agents can be seen as a distribution of points for which we're required to predict next destinations. Hence, even a classical machine learning approach could work.\n\nI'm a little bit surprised that none of available approaches goes in that direction 🤔 !?\n\nNevertheless, we're just at the beginning of this competition and there still, likely, lot of ideas to be discovered.\n\nAnd you, what's your experimental approach for now ?\n\n[Updates]\n[I finally train a LGBM model on a tabular version of the comeptition data. Feel free to comment with your ideas.](https://www.kaggle.com/kneroma/lgbm-on-lyft-tabular-data-inference)\n\nThe datasets used to train and predict with LGBM models are :\n* [train sample](https://www.kaggle.com/kneroma/lyft-motion-prediction-autonomous-vehicles-as-csv)\n* [test set as tabular data](https://www.kaggle.com/kneroma/lyft-test-set-as-csv)\n* [LGBM checkpoints](https://www.kaggle.com/kneroma/lyft-models)\n\n[Updates 2]\n* I've published my [training pipeline here](https://www.kaggle.com/kneroma/lgbm-on-lyft-tabular-data-training)\n* The whole training-set can be [found here as parquet file](https://www.kaggle.com/kneroma/lyft-train-as-parquet)", "1053639": "Hi,\n\nThanks @kneroma for the kernel. There's a [similar thread](https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/188330) where ideas like constant velocity Kalman filter models are discussed. The problem with using only the history positions is that they don't take into account the free space (other than those occupied by surrounding vehicles) and traffic light situation. That said, these methods can be used to smooth the output (from DL models) or can be used as one candidate in multiple predictions." }, "source": "meta" }