Improving Heuristics On-the-fly for Effective Search in Plan Space
21 Sep 2015-pp 302-308
TL;DR: This work extends single-step-error adaptation to heuristic functions from Partial Order Causal Link (POCL) planning to allow a partial order planner to observe the effective average- step-error during search.
Abstract: The design of domain independent heuristic functions often brings up experimental evidence that different heuristics perform well in different domains. A promising approach is to monitor and reduce the error associated with a given heuristic function even as the planner solves a problem. We extend this single-step-error adaptation to heuristic functions from Partial Order Causal Link (POCL) planning. The goal is to allow a partial order planner to observe the effective average-step-error during search. The preliminary evaluation shows that our approach improves the informativeness of the state-of-the-art heuristics. Our planner solves more problems by using the improved heuristics as compared to when it uses current heuristics in the selected domains.
Citations
More filters
Posted Content•
TL;DR: An online error minimization approach in POCL framework is discussed to minimize the step-error associated with the offline learned models thus enhancing their informativeness and scale up the performance of the planner over standard benchmarks, specially for larger problems.
Abstract: In recent years, the planning community has observed that techniques for learning heuristic functions have yielded improvements in performance. One approach is to use offline learning to learn predictive models from existing heuristics in a domain dependent manner. These learned models are deployed as new heuristic functions. The learned models can in turn be tuned online using a domain independent error correction approach to further enhance their informativeness. The online tuning approach is domain independent but instance specific, and contributes to improved performance for individual instances as planning proceeds. Consequently it is more effective in larger problems.
In this paper, we mention two approaches applicable in Partial Order Causal Link (POCL) Planning that is also known as Plan Space Planning. First, we endeavor to enhance the performance of a POCL planner by giving an algorithm for supervised learning. Second, we then discuss an online error minimization approach in POCL framework to minimize the step-error associated with the offline learned models thus enhancing their informativeness. Our evaluation shows that the learning approaches scale up the performance of the planner over standard benchmarks, specially for larger problems.
Cites methods from "Improving Heuristics On-the-fly for..."
...The online heuristic adjustment approach is inspired from a recent work presented as technical communication (Shekhar and Khemani 2015) tested in a few planning domains....
[...]
References
More filters
Proceedings Article•
13 Jul 2008TL;DR: A new technique is introduced for combining multiple heuristics in a single heuristic application, inspired by the evaluation functions used in two-player games, and can lead to a large reduction in the search effort at a small cost in the quality of the solution obtained.
Abstract: Heuristic functions for single-agent search applications estimate the cost of the optimal solution. When multiple heuristics exist, taking their maximum is an effective way to combine them. A new technique is introduced for combining multiple heuristic values. Inspired by the evaluation functions used in two-player games, the different heuristics in a single-agent application are treated as features of the problem domain. An ANN is used to combine these features into a single heuristic value. This idea has been implemented for the sliding-tile puzzle and the 4-peg Towers of Hanoi, two classic single-agent search domains. Experimental results show that this technique can lead to a large reduction in the search effort at a small cost in the quality of the solution obtained.
55 citations
Proceedings Article•
30 Jul 2000TL;DR: The planning graph structure that Graphplan builds in polynomial time, provides a rich substrate for deriving more effective heuristics for state space planners, and the mutex information in the planning graph captures exactly this interaction information.
Abstract: Graphplan and heuristic state space planners such as HSP-R and UNPOP are currently two of the most effective approaches for solving classical planning problems. These approaches have hither-to been seen as largely orthogonal. In this paper, we show that the planning graph structure that Graphplan builds in polynomial time, provides a rich substrate for deriving more effective heuristics for state space planners. Specifically, we show that the heuristics used by planners such as HSP-R and UNPOP do badly in several domains due to their failure to consider the interactions between subgoals, and that the mutex information in the planning graph captures exactly this interaction information. We develop several families of heuristics, some aimed at search speed and others at optimality of solutions. Our empirical studies show that our heuristics significantly out-perform the existing state space heuristics.
45 citations
Proceedings Article•
23 Apr 2002TL;DR: This paper addresses the role of ground actions in refinement planning, and presents empirical results indicating that their role is twofold: first, planning with ground actions represents a bias towards early commitment of parameter bindings, and second, ground actions help enforce joint parameter domain constraints.
Abstract: Less than a decade ago, the focus in refinement planning was on partial order planners using lifted actions. Today, the currently most successful refinement planners are all state space planners using ground actions—i.e. actions where all parameters have been substituted by objects. In this paper, we address the role of ground actions in refinement planning, and present empirical results indicating that their role is twofold. First, planning with ground actions represents a bias towards early commitment of parameter bindings. Second, ground actions help enforce joint parameter domain constraints. By implementing these two techniques in a least commitment planner such as UCPOP, together with using an informed heuristic function to guide the search for solutions, we show that we often need to generate far fewer plans than when planning with ground action, while the number of explored plans remains about the same. In some cases a vast reduction can also be achieved in the number of explored plans.
38 citations
01 Apr 1995
TL;DR: Improvements to plan refinement strategies for well-founded partial order planning are described that promise to bring this style of planning closer to practicality and propose an A* heuristic that counts only steps and open conditions, while ignoring "unsafe conditions".
Abstract: Describes some simple domain-independent improvements to plan refinement strategies for well-founded partial order planning that promise to bring this style of planning closer to practicality. One suggestion concerns the strategy for selecting plans for refinement among the current (incomplete) candidate plans. We propose an A* heuristic that counts only steps and open conditions, while ignoring "unsafe conditions" (threats). A second suggestion concerns the strategy for selecting open conditions (goals) to be established next in a selected incomplete plan. We propose a variant of a strategy suggested by Peot and Smith (1993) and studied by Joslin and Pollack (1994); the variant gives top priority to unmatchable open conditions (enabling the elimination of the plan), second-highest priority to goals that can only be achieved uniquely and otherwise uses LIFO (last-in, first-out) prioritization. The preference for uniquely achievable goals is a "zero-commitment" strategy in the sense that the corresponding plan refinements are a matter of deductive certainty, involving no guesswork. In experiments based on modifications of UCPOP (Unsafe Conditions Partial Order Planner), we have obtained improvements by factors ranging from 5 to more than 600 for a variety of problems that are nontrivial for the unmodified version. Crucially, the hardest problems give the greatest improvements.
27 citations
TL;DR: ‘‘Hybrid planning’’ refers to a distinct paradigm, which integrates building plans based on the individual action’s causal structure with obtaining plans from iteratively implementing abstract actions by pre-defined partial solutions.
Abstract: Planning and scheduling (P&S) constitute fundamental cognitive capabilities for systems to reason about plans and their causal structure. They are essential for producing a goal-oriented system behavior and supporting a user’s decision making. While planning is the method of creating courses of action that achieve goals or perform tasks, scheduling assigns consistent allocations of time and resources to activities. ‘‘Hybrid planning’’ refers to a distinct paradigm, which integrates building plans based on the individual action’s causal structure with obtaining plans from iteratively implementing abstract actions by pre-defined partial solutions. Combined, hierarchical model aspects represent regular solutions provided by domain experts, while causality-based techniques complete underspecified procedures and address exceptional cases. Applications of the technology range from planning out disaster relief missions [4] to assessing cyber-security measures [7].
22 citations