OPPORTUNISTIC CONTROL OF ACTION IN INTELLIGENT AGENTS

Given its multiple goals, limited resources, and dynamic environment, an intelligent agent must decide which of many possible actions to exe cute at each point in time. Planning and reactive models embody two di fferent modes of control. By contrast, we characterize a two-dimension al space of control modes, each of which maximizes the quality of run- time behavior in the corresponding region of a two-dimensional space o f control situations. The situation space is defined by dimensions rep resenting the predictability of the agent's task environment and the c onstraint imposed by its goals. The space of control modes is defined by corresponding dimensions representing the agent's sensitivity to ru n-time events and its commitment to specific actions prepared in advan ce. The corners of the mode space are anchored by extreme modes that a n agent rarely uses: open-loop plan execution, goal-specific reactivit y, dead reckoning, and reflex modes. The large interior is occupied by a space of more practical strategic control modes that modulate the b alance of sensitivity and commitment for situations embodying differen t combinations of predictability and constraint. To support behavior t hroughout the space of control modes, we propose an opportunistic cont rol model: run-time conditions trigger a subset of possible actions, s trategic plans constrain intended actions, and the match between possi ble actions and strategic plans controls action execution. The model i s operationalized as an agent architecture and an associated represent ation language for actions and plans. We show how the model allows an agent to move continuously throughout the space of control modes and r eport preliminary empirical results.