A profile of the Australian Artificial Intelligence Institute [World Impact]

architectures. To make the theory of BDI agents more practical, we need the notion of plans supplied in advance, rather than generated by the agent. For such agents, deliberation relates to selecting a plan. Plan execution consists of the hierarchical expansion of these plans, guided by a means-end analysis. Augmented with plans, we can provide an abstract architecture of BDI agents and relate it to implemented systems such as dMARS. Team-oriented systems. By extending the notion of single agents, we have investigated multiple agents working as a team and the associated theory of mutual beliefs, joint goals, and joint intentions. As a natural extension of plans, we have joint plans that act as recipes or coordination protocols for multiple agents. This is elaborated in our research on planned team a~t ivi ty .~ Reactive plan recognition. An agent’s recognition of the mental state (beliefs, desires, plans, and intentions) of other agents is an important part of intelligent activity. Doing this with limited resources and with a continuously changing environment is a challenge. We have extended the philosophy of using plans to this task, and call the approach reactive plan recognition. We provide algorithms for reactive plan recognition and embed them in the framework of agent-based reasoning. This results in a model for mental-state recognition with integrated reactive plan execution and recognition. This approach has been applied in an air-combat model to enable pilots to infer their opponents’ mental state and choose their tactics accordingly. Agent-oriented methodology. Extensive work with end users initiated a research program to enable software analysts and engineers, rather than researchers, to design, implement, and maintain BDI systems. We developed an agent-oriented methodology and modeling technique for systems of agents based on object-oriented models. By expanding existing techniques, we can produce an approach that those familiar with the obejct-oriented paradigm can easily learn and ~nderstand.~ Agent-oriented languages. We developed an agent-oriented language, Agentspeak, that captures the essential features of BDI agents. It can be viewed as a simplified, textual language of an agent-based system such as dMARS. Agentspeak is a programming language based on a restricted first-order language with events and actions. Unlike our other work, the agent’s beliefs, desires, and intentions are not explicitly represented. This shift in perspective from an external viewpoint is likely to have a better chance of unifying theory and practice. Experimentation. The AA11 performs experiments on constrained and welldefined domains to investigate how commitment to goals contributes to effective behavior. We also compare different strategies for reacting to change. The results demonstrate the feasibility of developing systems that empirically measure agent performance. The combination of c o m t ment with intelligent reactive replanning results in optimal behaviors.