1  # Machine learning to improve sim racing driving styles
  2  
  3  Sim racing is about improving lap times. There are tools that capture telemetry
  4  and let you compare your telemetry to other drivers.
  5  
  6  https://www.youtube.com/watch?v=sGCuP1VCkTU
  7  
  8  But how about automtically detecting the best way to improve your driving style
  9  and giving recommendations *while* you are driving?
 10  
 11  ## Driving style changes
 12  
 13  Implementation of a change should result in an improvement of lap time.
 14  Changes can be compared by the amount of lap time, or sector time, improvement.
 15  
 16  Time improvement should be measured over the course of one or multiple corners.
 17  
 18  If there is one corner between two straights, the time measurement would start
 19  at the braking zone and end at the acceleration zone. Or the corner entry and exit.
 20  One might factor entry speed and exit speed into the measurement, since it's not
 21  only the time spend driving through a corner, but the overall lap time what matters.
 22  
 23  A more complex scenario might be the combination of successive corners, such as a chicane,
 24  which is usually a left followed by a right, or vice versa.
 25  
 26  ## Braking point
 27  You brake too early or too late
 28  Input: track position percent, brake input
 29  
 30  ## Trail braking
 31  Instead of releasing the brake apprubtly, you ease off the brakes while turning the car into a corner.
 32  Input: rate of change of brake input
 33  
 34  ## Acceleration
 35  Instead of going full on the throttle, you ease into the throttle.
 36  Input: rate of change of throttle input
 37  
 38  ## Overlapping brake and throttle
 39  
 40  ## Grip
 41  Having the tires not at full grip, i.e. either you are too slow or turning to slow and the tires still have grip available or you are braking too hard or accelerating too much and the tires either block or lose grip.
 42  
 43  ## Racing line
 44  Not following the racing line
 45  Input: track position x,y
 46  
 47  
 48  ## Geometric approach
 49  
 50  See [analysis](analysis/) folder for a geometric approach to the problem.
 51  This also contains some analysis EDA and visualization for lap data.
 52  
 53  ## Reinforcement learning
 54  We have human driver telemetry data as input and car telemetry data and world positions as output.
 55  The function to be maximised is corner time, sector time and lap time.
 56  The agent inputs are the car telemetry data and the world positions.
 57  The agent outputs are human driver inputs to the car, such as throttle, brake and steering.
 58  The world reaction are the car telemetry data, such as tire grip and speed and the world positions, such as percent of track completed.
 59  
 60  
 61  Q: would the agent learn the best combination of all laps used to train the agent? I.e. would we get a better lap than a human selected coach lap?
 62  
 63  Q: would we get an agent model that would respond with the best next car input given a current world and car telemetry?
 64  
 65  
 66  
 67  ## AI Coaching
 68  
 69  The coach should provide the driver with actionable advice on how to improve the lap time.
 70  Therefore the coach needs to evaluate what changes have the biggest impact on the lap time.
 71  This could be posed as a ranking problem.
 72  
 73  Given a corner inputs, find a recorded corner that is different in only one feature and compare the time improvement.
 74  Do this for all available features and select the one with the biggest improvement.
 75  https://towardsdatascience.com/time-series-clustering-deriving-trends-and-archetypes-from-sequential-data-bb87783312b4
 76  
 77  1. identify which cluster you belong to
 78  2. find adjacent clusters
 79  
 80  
 81  It also needs to evaluate past behaviour and incorporate changes the driver has made.
 82  E.g. if you usually brake too early in your previous laps, but now you fixed that behaviour, the coach should not notify you about that anymore.
 83  So the coach needs to have some memory of some sorts.
 84  
 85  
 86  ## Braking coach
 87  
 88  Given a position on the track, the coach should notify the driver to brake in 100m, 50m and now.
 89  
 90  The model is trained to emit a braking point signal 100m, 50m and now for the given position.
 91  The model should have some memory for the previous inferences, since we dont know about the frequency of inference, we can only assume that the previous inference is position wise before the current inference.
 92  
 93  
 94  ## Resources
 95  
 96  https://arxiv.org/abs/2005.05178
 97      https://github.com/linklab-uva/deepracing
 98  https://paperswithcode.com/dataset/torcs
 99  
100  
101  https://jmotif.github.io/sax-vsm_site/morea/algorithm/SAX.html
102  https://en.wikipedia.org/wiki/Q-learning
103  
104  https://theoehrly.github.io/Fast-F1/
105  ## Offline reinforcement learning
106  * https://arxiv.org/abs/2005.01643
107  * https://arxiv.org/pdf/2203.01387.pdf
108  
109  
110  ## Explaining a race track
111  
112  https://www.formula1-dictionary.net/corners.html
113  
114  ## Extracing features
115  
116  [Information about iRacing tracks](https://github.com/SIMRacingApps/SIMRacingApps/issues/242)
117  
118  ## track processing
119  
120  * https://github.com/TUMFTM/racetrack-database
121  * https://github.com/TUMFTM/global_racetrajectory_optimization
122  
123  # plotly rendering
124  * https://github.com/plotly/plotly.py/issues/931#issuecomment-819502866