Skip to content

Latest commit

 

History

History
64 lines (47 loc) · 2.46 KB

README.md

File metadata and controls

64 lines (47 loc) · 2.46 KB
Raw

FormulaE-EnvisionRacing

Problem Statement

To build a machine learning model that predicts the Envision Racing drivers’ lap times for the all-important qualifying sessions that determine what position they start the race in. Winning races involves a combination of both a driver’s skills and data analytics. To help the team you’ll need to consider several factors that affect performance during a session, including weather, track conditions, and a driver’s familiarity with the track.

Data Description

Weather dataset contains the following data from start time to end time of each session at 1 min interval

  • Air Temp
  • Humidity
  • Pressure
  • Track Temp
  • Wind Speed
  • Wind Direction
  • Rain
  • Location
  • Event (Free Practice / Qualifying Session)

Train dataset contains Location 1-5 Free practice and Qualifying Events

  • Driver detailes (Driver Number, Name, Team, Group)
  • Sector times (S1,S2,S3 and S1_LARGE, S2_LARGE, S3_LARGE)
  • Sector improvements
  • Pit time
  • Elapsed time**
  • Power & Max Speed
  • Event
  • Location

Approach

  • Cleaned weather data as it contains corrupted data in track side variables
  • Cleaned corrupted team name in train data
  • Converted elapsed time into to minute index based on their sector times and merged it with the indexed weather data, we get weather available for each sector
  • Pit time are overlaped with either S1 of next lap or S3 of previous lap, cleaned the data to get proper sector times for each lap
  • Standardized the weather data as the each numeircal values are in different unit, replaced missing values with -999
  • Train data is further split into 80/20 for cross validation

Evalualtion Metrics

Submissions are evalualted based on the RMLSE values

Models

Catboost was the base model used

Best model - baseline-lights-out-and-away.ipynb - 0.7232

Exprerimented Approaches

  • Ran a H2O AutoML "GLM" model with just Sector and Pit times
  • Used Bayesian Optimization on Catboost for hyperparameter tuning
  • Considered the residual errors of the regression model and used it as a output variable for Catboost+Bayes
  • Ran a AutoML H2O model

Future Enhancement (TO-DO)

  • Run A Deeplearning RNN/LSTM model
  • Consider the lag values of previous lap
  • Create a timeline based visualization for each event and location to check each driver position