Using a genetic algorithm to learn behaviors for autonomous vehicles
Alan C. Schultz, John J. Grefenstette · Guidance, Navigation and Control Conference · 1992
Truly autonomous vehicles will require both projec-tive planning and reactive components in order to perform robustly. Projective components are needed for long-term planning and replanning where explicit reasoning about future states is required. Reactive components allow the system to always have some action available in real-time, and themselves can exhibit robust behavior, but lack the ability to expli-citly reason about future states over a long time period. This work addresses the problem of creating reactive components for autonomous vehicles. Creat-ing reactive behaviors (stimulus-response rules) is generally difficult, requiring the acquisition of much knowledge from domain experts, a problem referred to as the knowledge acquisition bottleneck. SAMUEL is a system that learns reactive behaviors for auto-nomous agents. SAMUEL learns these behaviors under simulation, automating the process of creating stimulus-response rules and therefore reducing the bottleneck. The learning algorithm was designed to learn useful behaviors from simulations of limited fidelity. Current work is investigating how well behaviors learned under simulation environments work in real world environments. In this paper, we describe SAMUEL, and describe behaviors that have been learned for simulated autonomous aircraft, auto-nomous underwater vehicles, and robots. These behaviors include dog fighting, missile evasion, track-ing, navigation, and obstacle avoidance. 1.