Gregg Collins

Planning for contingencies: a decision-based approach

A fundamental assumption made by classical AI planners is that there is no uncertainty in the world: the planner has full knowledge of the conditions under which the plan will be executed and the outcome of every action is fully predictable. These planners cannot therefore construct contingency plans, i.e., plans in which diffierent actions are performed in diffierent circumstances. In this paper we discuss some issues that arise in the representation and construction of contingency plans and describe Cassandra, a partial-order contingency planner. Cassandra uses explicit decision-steps that enable the agent executing the plan to decide which plan branch to follow. The decision-steps in a plan result in subgoals to acquire knowledge, which are planned for in the same way as any other subgoals. Cassandra thus distinguishes the process of gathering information from the process of making decisions. The explicit representation of decisions in Cassandra allows a coherent approach to the problems of contingent planning, and provides a solid base for extensions such as the use of diffierent decision-making procedures.

Opportunistic Planning and Freudian Slips

Freud’s study of the psychology of errors (see, e. g., Freud, 1935), including notably slips of the tongue, led him to the conclusion that many such errors are not merely the result of random malfunctions in mental processing, but rather are meaningful psychological acts. That is, they are intentional actions in every sense of the word, reflecting and indeed carrying out the goals, whether conscious or not, of the person who commits them. In particular, Freud argued, such errors stem from attempts to carry out suppressed intentions, intentions that have been formed but then in some sense withdrawn because they conflict with other, more powerful intentions.

The Role of Self-Models in Learning to Plan

We argue that in order to learn to plan effectively, an agent needs an explicit model of its own planning and plan execution processes. Given such a model, the agent can pinpoint the elements of these processes that are responsible for an observed failure to perform as expected, which in turn enables the formulation of a repair designed to ensure that similar failures do not occur in the future. We have constructed simple models of a number of important components of an intentional agent, including threat detection, execution scheduling, and projection, and applied them to learning within the context of competitive games such as chess and checkers.

An interface design tool based on explicit task models

Producing high-quality, comprehensible human interfaces is a difficult, labor-intensive process that requires experience and judgment. In this paper, we describe an approach to assisting this process by using explicit models of the user’s task to drive the interface design and to serve as a functional component of the interface itself. The task model helps to ensure that the resulting interface directly and transpru-ently supports the user in performing his task, and serves as a scaffolding for providing in-context help and advice. By crafting a library of standardized, reusable tasks and interface constructs, we believe it is possible to capture some of the design expertise and to amortize much of the labor required for building effective user interfaces.

Issues in the justification-based diagnosis of planning failures

Publisher Summary This chapter reviews the issues involved in the justification-based diagnosis of planning failures. A planner in the real world must constantly contend with uncertainty. Failure once in the run of plans as predicted by the planner is forgivable but a repetitive pattern of failures is intolerable. This argument constitutes the main motivation for a failure-driven approach to explanation-based learning in planning domains. A post hoc analysis of the failure(s), of which assumptions failed, and how and why they failed should be carried out automatically by the planner. To carry out such an analysis, the planner must first of all be able to determine the assumptions underlying its plan, and their role in the reasoning that led it to expect that the plan would perform as intended. The basic process of failure diagnosis, given such justification structures, is a simple recursive procedure. If an expectation fails, it is to be checked whether any of the antecedents of the rule that generated the expectation are contradicted by the facts of the case. If so, then the antecedent(s) that have been contradicted and recur have to be faulted. Otherwise, the problem lies in the assumptions associated with the rule. The chapter presents a test-bed system comprising mechanisms for inference and rule application, justification maintenance, expectation handlers, failure explanation, and rule patching.

Model-based integration of planning and learning

The goal of our research is to construct an integrated model of planning and learning that can account for the acquisition of new planning knowledge. Our approach involves the use of model-based reasoning. In this approach, the system monitors its performance by comparing it with expectations derived from a model of the systems planning architecture. The arguments relating the systems expectations to its underlying model of the planning process are encoded in the form of explicit justification structures. When the systems actual performance diverges from its expectations, it traces back through these justification structures, looking to fault the setting of some controllable parameter of the planner. When such a controllable parameter is isolated, a repair is then effected, in the form of an adjustment to one of these parameters.

Improving decision-making on the basis of experience

Publisher Summary This chapter discusses the improvement of decision making on the basis of experience. Decision making in general is subject to a trade-off between the sophistication of the procedure employed to make the decision, and the time and effort required to perform the analysis. Whether a more sophisticated procedure is worth using depends on whether the gain from its superior performance outweighs the extra cost. The key point about such trade-offs in the sophistication of the decision-making process is that, like the decisions themselves, they are highly dependent upon the particular planning environment. In the absence of a theory of the utility of decision-making sophistication as a function of the planning domain, the appropriate level of sophistication for a given domain must be determined empirically. The ability of a planner, in a decision making process, to adapt its decision-making to a domain depends in part upon its ability to optimize the trade-off between the sophistication of its decision procedures and their cost. As it is difficult to optimize this trade-off on a priori grounds alone, a planner should start with a relatively simple set of decision procedures and add complexity in response to experience gained in the application of its decision-making to real-world problems.