Prototyping Intelligent Vehicle Modules Using Evolutionary Algorithms

Abstract

Intelligent vehicles must make real-time tactical level decisions to drive in mixed traffic environments. SAPIENT is a reasoning system that combines high-level task goals with low-level sensor constraints to control simulated and (ultimately) real vehicles like the Carnegie Mellon Navlab robot vans. SAPIENT consists of a number of reasoning modules whose outputs are combined using a voting scheme. The behavior of these modules is directly dependent on a large number of parameters both internal and external to the modules. Without carefully setting these parameters, it is difficult to assess whether the reasoning modules can interact correctly; furthermore, selecting good values for these parameters manually is tedious and error-prone. We use an evolutionary algorithm, termed Population-Based Incremental Learning, to automatically set each module's parameters. This allows us to determine whether the combination of chosen modules is well suited for the desired task, enables the rapid integrationof new modules into existing SAPIENT configurations, and provides an automated way to find good parameter settings.