Learning to Race: Experiments with a Simulated Race Car

Our focus is on designing adaptable agents for highly dynamic environments. Wehave implementeda reinforcement learning architecture as the reactive componentof a twolayer control system for a simulated race car. Wefound that separating the layers has expedited gradually improving performance. Weran experiments to test the tuning, decompositionand coordination of the low level behaviors. Ourcontrol system wasthen extendedto allow passingof other cars and tested for its ability to avoidcollisions. Thebest designused reinforcement learning with separate networksfor each action, coarse codedinput and a simple rule based coordination mechanism.

• Type: Conference
• Year: 1998
• Where: FLAIRS
• Authors: Larry D. Pyeatt, Adele E. Howe
• Keywords: Artificial Intelligence, FLAIRS 1998, Reinforcement Learning, Reinforcement Learning Architecture, Simulated Race Car