"Learning AI" For 3D Car Racing
"Learning AI" For 3D Car Racing
Developing Platform
• The Open Racing Car Simulator (TORCS)
Client-server
Real-time execution
Physical seperation between server and the driver code
Objective
• Given a driving simulation envirnment, design
an AI bot which can drive.
• Train itself to drive better and better in a given
map.
• By driving on different maps, learn the ‘driving
trick’and be able to drive fairly good on a new
map.
• And more
Hypothetical Solutions
• Q1: Given a driving simulation envirnment, de
sign an AI bot which can drive.
• S: Looking for longest distance ahead.
Hypothetical Solutions
• Q2:Train itself to drive better and better in a gi
ven map.
• S:Tune the coef.(Max speed, steering coef. bra
ke etc.) by repeatly run.
Hypothetical Solutions
• Q3:By driving on different maps, learn the ‘dri
ving trick’and be able to drive fairly good on a
new map.
• S: Map analysis.
Execution & Chanllenges
• Big turn (e.g U-shape)
Execution & Chanllenges
• Large amount of trial run.
Steering x Max speed x Brake x Acc x....
Initiate
population Trial run Cross over
Selection Mutation
Execution & Chanllenges
• Neural Network