Post on 31-Mar-2015
Optimized Sensory-motor Couplings plus Strategy Extensions for the TORCS Car Racing Challenge
COBOSTAR Cognitive BodySpaces for Torcs based Adaptive
Racing
Outline:• TORCS Competition Setup
• COBOSTAR Design
• Parameter Optimization
• Strategy Modifications
TORCS Competition Setup
TORCS Competition Setup:
Only local information /
Main Idea
no complete track Information available
TORCS Competition Setup:
Available Sensors
• Angle Sensor: current angle between the car direction and the track axis.
• Speed Sensors: speed in both axes.• 19 Range Sensors: free track space in front of
the car.• 36 opponent Sensors: notice opponents around
the car.• Additional Sensors: current engine speed,
current gear, rotation and speed for all wheels, damage of the car, etc.
TORCS Competition Setup:
Car Control•Gas and Break Pedals•Gear shifting•Steering
COBOSTAR Design
COBOSTAR Design:
1 .On Track Strategy• Distance and direction of the longest range
sensor
2 .Off Track Strategy• Angle to track axis• Relative position to track center
COBOSTAR Design:
On Track Strategy
Using:
Calculate: Target steering
angle:
Target speed:
Finally:
COBOSTAR Design: on Track Strategy
:Gas or Break
How to Break :
How to Steer :
COBOSTAR Design: off Track Strategy
what changes?
• Distance sensors to track borders are
unavailable.
• Steering becomes more difficult.
• Wheel slippage is much stronger.
Using:
Calculate:Target steering angle:
Target speed:
COBOSTAR Design: off Track Strategy
Stuck behavior:
COBOSTAR Design: off Track Strategy
Anti Slip Regulation:
If nothing else works,
Switch to reverse gear and stay in this mode until
angle to track axis is halved or until stuck again
Parameter Optimization
Parameter Optimization
CMA – Covariance Matrix AdaptationSmart(er) search for the best solution. Takes in
account the dependencies between the parameters
Evolution Process:• All parameters were optimized on various TORCS
tracks.• All sets of parameters were compared on all
tracks.• The most general parameter set was chosen.
Optimization Algorithm:
Parameter Optimization – On Road
Fitness Function: 1( / distance raced + 1)
Parameter Optimization – On Road
• The most general set, wins only in four of the tracks
(not that general).
• Second set with most wins, is only rated 6th – different
behavioral strategies suite for different track types.
• Best strategy for a track isn’t always the strategy that
was optimized on that track – local optimum.
• Differences between worst and best performance on
each track – hard to get a general strategy
Interesting findings
Parameter Optimization – On Road
Blast from the past:
some strategies control this formula with P1 and some with P2.
Parameter Optimization – Off Road
• Not the same Fitness function. If the car controller is
good, then the car will not reach “off-road”
Why not the same as on-road?
• Crash Strategy: every 300 meters causes the
car to go off the road.
• Now the same fitness function can be used.
Solution
Strategy Modifications
Strategy Modifications – Gear Shifting
Shift up:
Shift down:
Each time the engine reaches 9500 RPM
Each time the engine drops bellow:
3300, 6200, 7000, 7300, 7700
For gears: 2, 3, 4, 5, 6 respectively
Strategy Modifications – Large Track
The problem:
Solution:
The target angle is interpolated between the
maximal distance sensor and its neighbors,
causing the to drive on slightly wavy trajectory.
Measuring the track width at the beginning of the race. If
it exceeded a hand-set threshold, some steering factors
were set by hand instead of the evolved ones.
Strategy Modifications – 2nd Lap
Switching Strategies:
Analyze the general properties of the track and
switch to more suitable strategy.
Crash Point:
Remembering crash points from the first lap, in
the next lap the car would go into “passive
mode”.
QUESTIONS?