Multi-objective planning using a metric sensitive planner

Automated planning addresses the problem of generating a sequence of actions to satisfy given goal conditions for a constructed model of the world. In recent planning approaches heuristic guidance is used to lead the search towards the goal. The focus of this work is on domains where plan quality is assessed with plan metrics. A discussion of the impact of a popular relaxed planning graph heuristic on the quality of plans in such domains is presented. The relaxed planning graph heuristic bias towards shorter plans, irrespective of quality, is described. A novel approach to constructing the relaxed planning graph based on metric cost is presented to overcome this bias and to generate good quality plans. A notion of metric sensitivity as the ability of a planner to respond to the change of the plan metric, is introduced and methods to determine metric sensitivity are presented. Current state-of-the-art planners are evaluated in terms of their metric sensitivity. This research also tackles the problem of planning in multiobjective domains, where quality of a plan is evaluated using multiple plan metrics. For multiobjective domains the solution is no longer a single plan but a set of plans. A set of non dominated solutions is called a pareto frontier. This thesis contains a discussion on the desired properties of such sets of plans and methods of generating them. Metric sensitivity is a required property for a planner to effectively reason with user defined metrics and generate desired set of plans. The main significant contributions of the work described in the thesis are: 1. A definition and exploration of metric sensitivity in planning. 2. A context-dependent, cost-based relaxed planning graph and heuristic. 3. A compilation method from cost to temporal domains. 4. Examination of the impact of planners’ properties on the quality of plans and APFs.