Multiple adaptive agents for tactical driving

摘要

Recent research in automated highway systems has ranged from low-level vision-based controllers to high-level route-guidance software. However, there is currently no system for tactical-level reasoning. Such a system should address tasks such as passing cars, making exits on time, and merging into a traffic stream. Many previous approaches have attempted to hand construct large rule-based systems which capture the interactions between multiple input sensors, dynamic and potentially conflicting subgoals, and changing roadway conditions. However, these systems are extremely difficult to design due to the large number of rules, the manual tuning of parameters within the rules, and the complex interactions between the rules. Our approach to this intermediate-level planning is a system which consists of a collection of autonomous agents, each of which specializes in a particular aspect of tactical driving. Each agent examines a subset of the intelligent vehicle's sensors and independently recommends driving decisions based on their local assessment of the tactical situation. This distributed framework allows different reasoning agents to be implemented using different algorithms. When using a collection of agents to solve a single task, it is vital to carefully consider the interactions between the agents. Since each reasoning object contains several internal parameters, manually finding values for these parameters while accounting for the agents' possible interactions is a tedious and error-prone task. In our system, these parameters, and the system's overall dependence on each agent, is automatically tuned using a novel evolutionary optimization strategy, termed Population-Based Incremental Learning (PBIL). Our system, which employs multiple automatically trained agents, can competently drive a vehicle, both in terms of the user-defined evaluation metric, and as measured by their behavior on several driving situations culled from real-life experience. In this article, we describe a method for multiple agent integration which is applied to the automated highway system domain. However, it also generalizes to many complex robotics tasks where multiple interacting modules must simultaneously be configured without individual module feedback. [References: 24]