Real-Time Agents: Reaction vs. Deliberation 1

In agents theory, it is commonly accepted that reactiv- ity is one of the main features of an agent. Reactivity can be defined as the capability of an agent to respond to significant changes in its environment. Traditionally, reactivity has been confronted with the agent's capability of deliberation, in the sense that the most reac- tive an agent is, the least time it spends deliberating (and vice-versa). Agent architectures normally present a fixed proportion between re- action and deliberation, normally implemented by assigning a given amount of resources to each of them at design time, with no possibil- ity of further adaptation at run time. In this way, the agent may work well for certain environments/problems, but it can poorly adapt this feature to changes in such initial conditions. Therefore, if the agent could accommodate its reactivity to the cur- rent situation of the environment, its adaptability would be consider- ably enhanced and its behavior would be closer to humans. Further- more, if the agent has real-time requirements, the agent's ability to adapt its reactivity becomes essential, because the environment will typically undergo periods of different stress conditions. In this sense, this paper introduces the concept of Reactivity Degree. This concept implies some meta-reasoning capabilities to be available in the agent, in order to dynamically decide the amount of resources which have to be assigned to deliberation and reaction. The paper also shows how to implement such concept in a hard real-time, hybrid agent archi- tecture named ARTIS, as well as some experimental results which demonstrate the usefulness of this new concept.