BDI-Dojo: developing robust BDI agents in evolving adversarial environments

Abstract

The Belief-Desire-Intention (BDI) architecture is a widely-used model for developing multi-agent systems. BDI agents pursue their goals over time using a collection of plan recipes that are programmed by the developers. Thus, traditional BDI agents are limited in dealing with dynamic environments where uncertainties are not known beforehand, such as those introduced by adversarial forces. In this paper, we present the BDI-Dojo framework for developing robust BDI agents by training them using reinforcement learning against similarly learning-equipped adversarial agents. This adversarial training approach empowers BDI agents to become more resilient in uncertain, dynamic environments.