the 2011 simulated car racing championship @ evo*-2011

2011 Simulated Car Racing Championship @ EVO*-2011

The 2011 Simulated Car Racing

Championship @ Evo*-2011Daniele Loiacono, Luigi Cardamone,

Martin V. Butz, and Pier Luca Lanzi


2011 Simulated Car Racing Championship9 races during 3 conferences

Develop a driver for TORCS(hand-coded, learned, evolved, …)

Drivers will be awarded based on their score

in each conference competition

At the end, the team with highest overall scorewins the championship

2


Roadmap of the 2011 Championship

1. EVO*-2011, Torino (Italy)April 27-29, 2011 (Now!)

2. ACM GECCO-2011, Dublin (Ireland)July 12-16, 2011

3. IEEE CIG-2011, Seoul (South Korea)August 31 September 3, 2011


Motivations

Proposing a relevant game-based competition

more representative of commercial games AI

more similar to a real-world problem

Proposing a funny and exciting competition

you can see and play with the entries of this competition

human players can interact with AI

a lot of programmed AI available for comparison

Proposing a challenging competition

computationally expensive

real-time

on-line learning

noisy sensors

The Open Racing Car Simulator


The Open Racing Car Simulator

TORCS is a state of the art open source simulator written in C++

Main features

Sophisticated dynamics

Provided with several cars, tracks, and controllers

Active community of users and developers

Easy to develop your own controller

OS Support

Linux: binaries and building from sources

Windows: binaries and “limited” bulding from sources support

OSX: legacy binaries and no building from sources support


The Open Racing Car Simulator & the Competition Software

TORCS

BOT BOT BOT

TORCS

PATCH

SBOT SBOT SBOT

BOT BOTBOT

UDP UDPUDP The competition server

Separates the bots from TORCS

Build a well-defined sensor model

Works in real-time


Sensors and actuators

Rangefinders for edges on the track and opponents (with noise)

Speed, RPM, fuel, damage, angle with track, distance race, position on track, etc.

Six effectors: steering wheel [-1,+1], gas pedal [0, +1], brake pedal [0,+1], gearbox {-1,0,1,2,3,4,5,6}, clutch [0,+1], focus direction

Structure of the Race


What is the structure of a race?

Three stages: warm up, qualifiers, actual race

During warm-up, each driver can explore the track and learn something useful

During qualifiers, each driver races alone against the clock (the best 8 drivers move to the race)

During the race all the drivers race together

12

Competitors


The competitors

Four entries in the first leg:

Mr. Racer, TU Dortmund

Mariscal & Fernández, Málaga University

Ready2Win, Slovak University of Technology - FIIT

Powaah, Blekinge Institute of Technology

Three more entries from the 2009 championship:

AUTOPIA, Madrid and Granada

COBOSTAR (Lönneker & Butz, University of Würzburg)

POLIMI (Cardamone, Politecnico di Milano)


Jan Quadflieg, Tim Delbruegger and Mike Preuss TU Dortmund

Mr Racer


Mr. Racer 2011

Mostly described in the Evo* Paper “Driving Faster Than a Human Player“

13 Variables learned offline with the CMA-ES

Warmup used to learn a model of the track

Noise Handling:

Low pass filter on sensor values

2 Regression polynoms are fitted to each side

Resample the polygons


Mr. Racer 2011 - Opponent Handling

Noise filtering with a lowpass filter

Observer module interprets filtered sensor readings

This leads to a recommended speed and an overtaking line (created on the fly!)

Planning module incorporates recommended target speed and overtaking line into the plan

Lots of nummerical parameters not yet tuned:

left for Gecco Competition


Mariscal & Fernández

General approach: expert systems improved by multi-objectives evolutionary computation techniques

Steering: based on a simple rule follow the track sensor with biggest distance.


Mariscal & Fernández - Throttle/Brake

Implemented via a parameterized expert system whoseparameters were optimized via a Multi-ObjectiveEvolutionary Algorithm (NSGA-II)

Two objectives:

Maximize distance.

Minimize damage.

Linear function (Ci are parameters to optimize):

C1 (Speed) + C2 (Distance to the next turn) + C3 (Turn degrees)

We obtain three sets of parameters

(i.e., controllers) to be used under

specific conditions of the race

(1) To run fast (2) To run safe(3) To run on dirtytracks.


Mariscal & Fernández - Overtaking

Expert system with four groups of rules:

Other strategies are implemented (via expert

systems) such as collision avoidance and slipstream overtakings.


Ready2Win

Modular architecture:

Driving module:

• The speed is computed using the maximum braking distance

Overtaking module

Recovery module

ABS module

Track learning during the warm-up:

First lap, driven slow, to identify turns (start, end, entry position, curvature) and learn the track model

Other laps, speed adaptation:

• Increasing according to lateral speed

• Reducing when the car goes off-road


Tim Uusitalo Blekinge Institute of Technology

Supervisor: Stefan Johanson

Powaah


Powaah

Based on Artificial Potential Field

Every sensor in the span (-40, +40) has a charge

The charge is determined by how far the sensor measures the track

Coulomb’s law is used to calculate potentials in the field

Pot = (C1 * C2 ) / d2

The sensor with the highest potential is used for computing how much steering is needed


Industrial Computer Science Department.

Centro de Automática y Robótica

Consejo Superior de Investigaciones Científicas

Madrid, Spain

Contact:E. Onieva ([email protected])

AUTOPIA


AUTOPIA

Fuzzy Architecture based on three basic modules for gear, steering and speed control

optimized with a genetic algorithm

Learning in the warm-up stage:

Maintain a vector with as many real values as tracklengthin meters.

Vector initialized to 1.0

If the vehicle goes out of the track or suffers damage then multiply vector positions from 250 meters before the current position by 0.95.

During the race the vector is multiplied by F to make the driver more cautious in function of the damage:

F=1-0.02*round(damage/1000)


COBOSTAR

Thies Lönneker and Martin V. Butz

University of Würzburg

http://www.coboslab.psychologie.uni-wuerzburg.de


COBOSTAR

One of the best controller of the 2009 Championship

Parameterized controller optimized with CMA-ES

Dynamically saves crash points


CIG-2008 Champ

Luigi Cardamone

Politecnico di Milano


Luigi Cardamone

Winner of the CIG-2008 Car Simulated Competition

The controller is based on a neural network evolved with NEAT

Very simple policy for overtaking

Qualifying


Scoring process: Warm-up & Qualifying

Scoring process involves three tracks:

OldPeak

Polimi

HillRoad

The tracks are not distributed with TORCS:

Generated applying evolutionary computation

The competitors cannot know the tracks

Each controller raced for 100000 game ticks in the warm-up stage and then its performance is computed in the qualifying stage as the distance covered within 10000 game ticks

Dirt-1

OldPeak


Qualifying: OldPeak

7441,21

4606,09

8159,67

5062,1

6799,89

7372,23

8202,92

0 1000 2000 3000 4000 5000 6000 7000 8000 9000

Ready2Win

Powaah

Mr. Racer

Mariscal&Fernandez

Cobostar

Cardamone

Autopia

Dirt-1

Polimi


Qualifying: Polimi

6343,6

3891,01

8570,81

7775,64

6694,85

96,1469

8307,57

0 1000 2000 3000 4000 5000 6000 7000 8000 9000

Ready2Win

Powaah

Mr. Racer

Mariscal&Fernandez

Cobostar

Cardamone

Autopia

Dirt-1

HillRoad


Qualifying: HillRoad

7528,29

10075,3

10250,3

10547,8

10006,7

8937,33

11088,4

0 2000 4000 6000 8000 10000 12000

Ready2Win

Powaah

Mr. Racer

Mariscal&Fernandez

Cobostar

Cardamone

Autopia


Qualifying summary

Driver OldPeak Polimi HillRoad Total

Autopia 10 8 10 28

Mr. Racer 8 10 6 24

Mariscal&Fernandez 3 6 8 17

Cobostar 4 5 4 13

Ready2Win 6 4 2 12

Cardamone 5 2 3 10

Powaah 2 3 5 10


How much does noise affect the performance?


HillRoad with and without noise

0 2000 4000 6000 8000 10000 12000

Ready2Win

Powaah

Mr. Racer

Mariscal&Fernandez

Cobostar

Cardamone

Autopia

NO-NOISY NOISY

The Race


Three Tracks

For each track we run 7 races with random starting grids

Each race is scored using the F1 point system (10 to first, 8 to second, 6 to third, …)

Two points to the controller with lesser damage

Two points for the fastest lap of the race

46


Race: OldPeak

Competitor Score

Autopia 12

Mr. Racer 8

COBOSTAR 6

Cardamone 5

Ready2Win 5

Mariscal&Fer

nandez 4

Powahh 2


Race: Polimi

Competitor Score

Autopia 10

Mariscal&Fer

nandez 8

Mr. Racer 8

COBOSTAR 7

Cardamone 4

Powahh 3

Ready2Win 3


Race: HillRoad

Competitor Score

Autopia 12

Mariscal&Fer

nandez 8

Mr. Racer 6

Powahh 5

COBOSTAR 4

Cardamone 3

Ready2Win 2


Final Results

Competitor OldPeak Polimi HillRoad Total

Autopia 12 10 12 34

Mr. Racer 8 8 6 22

Mariscal&Fernandez 4 8 8 20

COBOSTAR 6 7 4 17

Cardamone 5 4 3 12

Powaah 2 3 5 10

Ready2Win 5 3 2 10

What about the race results?

Race and Qualifying have similar outcomes(with the exception of Ready2Win)

AUTOPIA still very competitive

Opponent manegement is crucial on race start. Very difficult with noise!

Thank you!

Looking forward to see your controllers @ Gecco-2011