In this paper, we consider the issue of designing agents that successfully balance the amount of time spent in reconsidering their intentions against the amount of time spent acting to achieve them. Following a brief review of the various ways in which this problem has previously been analysed, we motivate and introduce a simple formal model of agents, which is closely related to the well-known belief-desire-intention model. In this model, an agent is explicitly equipped with mechanisms for deliberation and action selection, as well as a meta-level control function, which allows the agent to choose between deliberation and action. Using the formal model, we define what it means for an agent to be optimal with respect to a task environment, and explore how various properties of an agent's task environment can impose certain requirements on its deliberation and meta-level control components. We then show how the model can capture a number of interesting practical reasoning scenarios, and illustrate how our notion of meta-level control can easily be extended to encompass higher-order meta-level reasoning. We conclude with a discussion and pointers to future work.
展开▼
