Planning and Goal Interaction: The Use of Past Solutions in Present Situations

This payer presents WOK, a cased-based planner that makes use of memory structures based on goal interactions. WOK generates original plans, (which take the form of recipes), in the domain of Szechuan cooking, by modifying existing plans that are stored and then retrieved on the basis of the goal interactions that they deal with. The paper suggests an organization and indexing strategy that allows the retrieval and use of plans that overarch sets of goals rather than just individual goal situations. It also demonstrates how episodic knowledge can be used to guide planning and avoid past failures.* 1 Introductiou Anyone can plan At this point in time, nearly anyone in Artificial Intelligence can write a planner. The basic ideas exist as part of the knowledge we have as a field. The ideas that goals have plans associated with them; that some plans are stored as sets of coordinated steps or subplans; and that the achievement of most goals is in fact the achievement of many sub-goals all exist as givens in the field. What does not exist is the knowledge of how to deal with planning situations in which multiple interacting goals have to be planned for. While much work has been done in the identification of goal interactions such as CONFLICT, COMPETITION and CONCORD, [7], little to date has been done to incorporate these ideas into the knowledge system of a planner. The following then is a description of WOK (Well Organized Knowledge), a planner that makes use of goal interaction knowledge to categorize and organize plans in terms of the interactions that they deal with. WOK generates original plans in the domain of Szechuan cooking, by modifying existing plans that are stored and then retrieved on the basis of goal interactions specific to the domain (e.g. CONFLICTING TASTES:ONE-DOMINATES and CONTRASTING TASTES:E&UAL-FOOTING) rather than on the basis of specific contextual goals that the plans meet (e.g. the goal to include chicken or the goal to avoid hot tastes). WOK’s processing is guided by the top down application of its memories of past successes and failures, stored in terms of the goal interactions that they deal with. 2 Planners and Plan Interaction The problem of planning has been worked on since the earliest days of AI ( [3], [l], [4]), b u most of the stress has been on systems t that plan for a single goal. Such systems are concerned primarily