An expedited BDI agent architecture: Improving the responsiveness of agent-based autonomous systems for handling critical situations

Abstract

This work presents improvements on BDI agent architectures towards expedited behaviour. Standard and even real-time BDI architectures, due the characteristics of their deliberative processes, can take considerable time deciding what to do, and this can present a significant delay when a prompt reaction is required in a critical situation. To address this issue without simply adding a pure reaction layer, we introduce a novel, expedited, BDI architecture capable of maintaining effective reasoning while providing adequate and fast reaction to perceptions that occur in a critical situation (E2BA). By ‘adequate’ we mean that a proper action is promptly decided and that the agent temporarily enters into an exceptional operation mode. In this paper we present a concrete implementation for our proposal as a variation of Jason, the leading BDI programming framework. Considering that our main target applications lay in the domain of robotics, we also make our implementation suitable to be used in conjunction with the Robot Operating System (ROS). We evaluate the proposed mechanism through two experiments. The first experiment shows significant (at least 2.7x) reaction-time improvements obtained from using our expedited Jason in comparison with ‘standard’ Jason. Moreover, the results show that expedited Jason reaction-times are not significantly affected as the agent gets busier (it is constant), in contrast to the exponentially increasing response-time of standard Jason. The second experiment addresses the usage of the proposed architecture within a realistic software-in-the-loop (SIL) application scenario: controlling an uncrewed aerial vehicle (UAV) on a fire-fighting mission. This second experiment is not only important for its realistic nature and for reinforcing the reaction-time improvements, but also because it illustrates the importance of changing the agent’s operation mode when performing a failsafe procedure. The permanent overhead introduced by EB2A is very small, not more than 0.5% of the standard BDI reasoning cycle time.