On Becoming Reactive

Abstract

This chapter describes an autonomous robot agent called Theo-Agent that initially constructs explicit plans to solve new problems in its domain and converges to an agent that reacts directly to the features of its environment with an appropriate action. As the rules take into account the agent's multiple goals, why the agent has chosen to attend to its current goal, conditions under which the plan will satisfy the current goal, and the necessity of the action to achieving the agent's goal, they produce reactive, opportunistic decisions equivalent to those that would be produced by continually re-invoking costly planning methods. Theo-Agent has been used to learn such rules for a simulated blocks-world robot, a simulated thermostat, and an operating mobile robot. Theo-Agent is defined in terms of a frame whose slots and subslots define the agent's beliefs, or internal state. The fundamental cycle of Theo-Agent is to repeatedly access the value of its chosen.action slot and perform the action described therein. As its world is only partially observable, there are cases in which Theo-Agent cannot construct guaranteed plans.