Finally the sources of the controllers submitted at the Simulated Car Racing Competition at WCCI-2008 are available:
Computational Intelligence in Games
As you may know the results of the Car Racing Competition have been presented at the IEEE World Congress on Computational Intelligence 2008.
The final stage of the competition consisted in running a race with all the controller together. As in the first stage the scoring process involved 3 tracks: Ruddskogen, Street-1, and D-Speedway.
For each of the three selected tracks, we run ten races using a random permutation for the starting order of the controllers. At the end of each race, consisting of 3 laps, the performance is computed as the median (the 50th percentile) of the controllers score assigned on the basis of the F1 point systems.
The following table reports the summary of the competition results:
Then we performed an additional scoring session to compare the controllers submitted with the two programmed controllers provided to the competitors with the API of the competition and a controller evolved using NEAT by Luigi Cardamone (a student of Politecnico di Milano). The scoring has been performed in the same way except for the number of races run that were 5 instead of 10.
In the following video you can see the highlights of the competition.
Dear WCCI-2008 Car Racers,
The first stage of WCCI-2008 car racing competition is now completed. This first step involved 3 tracks and 8 controllers. The three tracks selected for testing the controllers are Ruddskogen, Street-1, and D-Speedway. Five controllers were provided by the participants and they are the actual competitors in the race. Other three controllers were provided as examples by the organizers: one was evolved by Luigi Cardamone, a student of the Politecnico di Milano, one by Julian Togelius, and one from Daniele Loiacono.
For each of the three selected tracks, we run each controller ten times. The performance has been computed as the median (the 50th percentile) over the ten runs to avoid any issue about skewness.
For each track, we sorted the performances and we used the F1 scoring system to assign a score to each controller (10 points to the 1st, 8 points to the 2nd, 6 points to the 3rd and 5/4/3/2/1 points to the other ones).
The score of a controller is computed the sum of the points that the controller received on the three tracks. The scoreboard of this first phase is reported below.
Note that this scoreboard does not decide the winner. The winner will be decided by a set of ten races involving all the five controllers which will be run tomorrow.
We wish to thank you all for participating to the competition. Stay tuned for the final scoreboard!
Meanwhile, we look forward to meet you at the workshop where movies will be shown and the results of the competition involving also the other three controllers will be shown.
Daniele, Julian, & Pier Luca
Luigi Cardamone, a master student of Politecnico di Milano, has just submitted a controller. The solution has been evolved using the Kenneth Stanley’s rtNEAT and the C++ API for TORCS provided in the competition package.
The final score of this controller is slightly lower than the one submitted by Matt Simmerson.
You can find the updated scoreboard on the homepage.