Opportunities and planning in an unpredictable world

An agent operating in an unpredictable world cannot simply construct plans ahead of time and expect them to work perfectly. Instead, it must adapt its plans to the circumstances that it actually encounters. To do this, it must acquire information about the actual state of the world, and recognize when unpredicted situations affect its goals. Both problems are addressed in this thesis. The requirement for information arises out of the need to make decisions, and is thus a goal-directed activity. Plans to achieve information goals can be constructed in the same way as plans for other goals. This thesis describes Cassandra, a contingency planner whose plans include provisions for explicit decisions and the resulting information acquisition; and PARETO, a plan execution system whose plans also make explicit provision for information acquisition arising out of the need to make decisions. Plan adaptation during execution requires a unifying framework within which the plan construction and plan execution processes can be integrated. In this thesis the argument is made that the consideration of opportunities provides such a framework, and a mechanism is presented that enables an agent to recognize unexpected opportunities on the fly and respond to them appropriately and in a timely manner. Its implementation in PARETO is described. This mechanism is based on reference features, features that are both cheap and functional. These features appear to be prevalent in everyday life. Reference features form the basis of a powerful focusing mechanism that can help an agent balance deliberation and action by indicating both when deliberation would be productive and what the deliberation should be about. Their use by PARETO in recognizing and reasoning about opportunities is described.