抄録
The ability to decide proper actions reactively and robustly in the partially unknown and uncertain environment is very essential to the highly autonomous spacecraft such as planetary rovers. The concept of behavior-based architecture or situated agent, which has received much attention in the field of autonomous agents recent years, is very attractive and promising for this purpose. One of the significant issues in this methodology is how to prepare the large set of reactive behavior rules for the agent. To deal with this problem, we propose an inductive behavior learning method, with which the agents can acquire their proper reactive behavior rules from their experiences autonomously. In this method, learned behavior rules are represented by a set of “behavior subgoals” - overlapping regions in the “attribute space” (sensor space) and the decision making is performed reactively by matching the sensor inputs and the behavior history with the acquired behavior subgoals. As a consequence, the difficulty in the construction of reactive behavior modules in autonomous agents is relieved, and the robustness against the inevitable uncertainty in the real world is increased. We also present some simulation results of the goal-pursuing / obstacle avoidance task performed by a planetary rover to show the effectiveness of the proposed method.
