Domain Independent Approaches for Finding Diverse Plans Biplav Srivastava IBM India Research Lab sbiplav@in.ibm.com Subbarao Kambhampati Arizona State University rao@asu.edu Tuan A. Nguyen University of Natural Sciences natuan@fit.hcmuns.edu.vn Minh Binh Do Palo Alto Research Center minhdo@parc.com Alfonso Gerevini University of Brescia gerevini@ing.unibs.it Ivan Serina University of Brescia serina@ing.unibs.it IJCAI 2007, Hyderabad, India (6 Authors from 3 continents, 4 countries, 5 institutions) Jan 09, 2007 Domain Independent Approaches for Finding Diverse Plans 1 Motivation Traditionally, Planning has been seen as a problem of finding a single plan for going from an initial to a goal state Often, we need a set of inter-related plans instead of a single plan C={c1,c2,…c} X={x1,x2,…x} I={i1, i2,… i} FPC Specifications Logical Physical Composition Composition RAW S={S1,S2,…SK} Jan 09, 2007 FRE Runtime RIW REW T= {t1,t2,…t} W={W1,W2,…WL} Domain Independent Approaches for Finding Diverse Plans 2 Motivation Traditionally, Planning has been seen as a problem of finding a single plan for going from an initial to a goal state Often, we need a set of inter-related plans instead of a single plan Diverse plans Jan 09, 2007 A set of web service compositions that can cover as much of the runtime failure circumstances as possible Or a set of intrusion plans that are qualitatively different Similar plans: plan stability (Fox et al ICAPS 06); a set of query plans so that partial results of time-out queries can be used First diverse, then similar; etc … We explore domain-independent approaches for finding diverse plans Domain Independent Approaches for Finding Diverse Plans 3 Finding Diverse plans How do we formulate and solve this problem? Naïve idea: Let the planner just continue to search for more plans It is not enough for the planner to just produce multiple plans. We want the plans to have some guaranteed diversity Domain-dependent approach Have a meta-theory of the domain in terms of predefined attributes and their possible values covering roles, features and measures. Use these attributes to compare plans [Myers ICAPS 2006] Issue: We are interested in domain-independent approach. Need to: Formalize notions of diversity (distance measures) Need to develop (or adapt existing) planning algorithms to search for diverse plans Jan 09, 2007 Needs extensive domain modeling Not affordable for many types of applications What bases for comparison are easier to enforce than others? How scalable are the algorithms? Domain Independent Approaches for Finding Diverse Plans 4 Outline Motivation Problem Formulation (s) Distance Measures Solution Approaches Jan 09, 2007 Different bases for comparison Different bases for computation Constraint-satisfaction based Heuristic-search based Results Related Work Conclusion Future Work Domain Independent Approaches for Finding Diverse Plans 5 Problem Formulation dDISTANTkSET Variations on the formulations possible Jan 09, 2007 Given a distance measure d(.,.), and a parameter k, find k plans for solving the problem that have guaranteed minimum pairwise distance d among them in terms of d(.,.) Converse formulation for dCLOSEkSET Related work – Multiple solutions for CSP problems (See Hebrard 2005, 2006) Domain Independent Approaches for Finding Diverse Plans 6 Distance Measures In what terms should we measure distances between two plans? Choice may depend on Jan 09, 2007 The actions that are used in the plan? The behaviors exhibited by the plans? The roles played by the actions in the plan? The ultimate use of the plans E.g. Should a plan P and a non-minimal variant of P be considered similar or different? What is the source of plans and how much is accessible? E.g. do we have access to domain theory or just action names? Domain Independent Approaches for Finding Diverse Plans 7 Basis for Comparing Plans Jan 09, 2007 Actions in the plan States in the behavior of the plan Causal support structures in the plan Domain Independent Approaches for Finding Diverse Plans 8 Quantifying Distances Set-difference Neighborhood based Jan 09, 2007 Prefix-based Suffix-based … Domain Independent Approaches for Finding Diverse Plans 9 Goal State Initial State Action Preconditions Effect A1 p1 g1 A2 p2 g2 A2’ p2, g1 g2 A1 <g1,g2,g3> p1, p2, p3 A3 <p1,p2,p3> A3 p3 g3 A3’ p3, g2 g3 Plan Plan S1-1 Goal Causal Chains p1, p2, p3 A1 <g1,p2,p3> S1-1, S1-2 S1-3 g1, g2, g3 A2 g1 Ai-p1-A1-g1-Ag g2 Ai-p2-A2-g2-Ag g3 Ai-p3-A3-g3-Ag g1 Ai-p1-A1-g1-Ag g2 Ai-p1-A1-g1-A2’,Ai-p2-A2’, A2’-g2-Ag g3 Ai-p3-A3’, Ai-p1-A1-g1-A2’,Aip2-A2’-g2-A3’, A3’-g3-Ag A2 A3 <g1,g2,p3> g1, g2, g3 <g1,g2,g3> <p1,p2,p3> Plan S1-2 p1, p2, p3 A1 <g1,p2,p3> A2’ A3’ <g1,g2,p3> g1, g2, g3 <g1,g2,g3> <p1,p2,p3> Jan 09, 2007 Plan S1-3 10 Compute by Set-difference Goal State Initial State A1 <g1,g2,g3> •Action-based comparison: S1-1, S1-2 are similar, both dissimilar to S1-3; with another basis for computation, all can be seen as different •State-based comparison: S1-1 different from S1-2 and S1-3; S1-2 and S1-3 are similar •Causal-link comparison: S1-1 and S1-2 are similar, both diverse from S1-3 p1, p2, p3 g1, g2, g3 A2 A3 <p1,p2,p3> Plan S1-1 p1, p2, p3 A1 <g1,p2,p3> A2 A3 <g1,g2,p3> g1, g2, g3 <g1,g2,g3> <p1,p2,p3> Plan S1-2 p1, p2, p3 A1 <g1,p2,p3> A2’ A3’ <g1,g2,p3> <g1,g2,g3> <p1,p2,p3> Plan S1-3 g1, g2, g3 Solution Approaches Possible approaches [Parallel] Search simultaneously for k solutions which are bounded by given distance d [Greedy] Search solutions one after another with each solution constraining subsequent search Explored in CSP-based GP-CSP classical planner Heuristic-based LPG metric-temporal planner Jan 09, 2007 Relative ease of enforcing diversity with different bases for distance functions Scalability of proposed solutions Domain Independent Approaches for Finding Diverse Plans 12 GP-CSP Result: Solving time with different bases Average solving time (in seconds) to find a plan using greedy (first 3 rows) and by random (last row) approaches Solving for diversity guided by distance functions is more efficient than random search Jan 09, 2007 Domain Independent Approaches for Finding Diverse Plans 13 GP-CSP Result: Solution quality time with different bases Comparison of the diversity in the solution sets returned by the random and distance function-guided greedy approaches Solving for diversity guided by distance functions is likely to get better quality of results than random search Jan 09, 2007 Domain Independent Approaches for Finding Diverse Plans 14 GP-CSP Result: Using different distance bases (time) Solving for diversity guided by dc or ds is easier (gives more results in the same time) than da Jan 09, 2007 Domain Independent Approaches for Finding Diverse Plans 15 GP-CSP Result: Using different distance bases (cross-validation on solution quality) Cell <row, column> = d’, d” indicates that over all combinations of (d,k) solved for distance d, the average value d”/d’ where d” and d’ are distance measured according to d” and d’ respectively. Example: <ds ,da> = 0.485 means that over 462 combinations of (d,k) solvable for ds for each d, the average distance between k solutions measured by da is 0.485 * ds. The results indicate that when we enforce d for da, we will likely find even more diverse solution sets according to ds (1.26* da) and dc (1.98* da ) Jan 09, 2007 Domain Independent Approaches for Finding Diverse Plans 16 Exploring with LPG • Details of changes to LPG in the paper • Looking for: • How large a problem can be solved easily • Large sets of diverse plans in complex domains can be found relatively easily • Impact of • = 3 gives better results • Can randomization mechanisms in LPG give better result? • Distance measure needed to get diversity effectively Jan 09, 2007 Domain Independent Approaches for Finding Diverse Plans 17 Experiments with LPG LPG-d solves 109 comb. Avg. time = 162.8 sec Avg. distance = 0.68 Includes d<0.4,k=10; d=0.95,k=2 Jan 09, 2007 LPG-d solves 211 comb. Avg. time = 12.1 sec Avg. distance = 0.69 Domain Independent Approaches for Finding Diverse Plans LPG-d solves 225 comb. Avg. time = 64.1 sec Avg. distance = 0.88 18 Related Work The problem of returning diverse relevant results is important in Information Retrieval The problem of finding “similar” plans has been investigated in Replanning and Plan Reuse. But limited notions of distance measures Myers 2006 gives a meta-theoretic basis for plan comparison For CSPs, Hebrard et al 2005 have formulated the problem and proposed solutions Jan 09, 2007 Think “relevance” “solution ness” The worst-case complexity results can be borrowed for planning Domain Independent Approaches for Finding Diverse Plans 19 Conclusion Contributions Formalize notions of bases for plan distance measures Proposed adaptation to existing representative, state-of-the-art, planning algorithms to search for diverse plans Jan 09, 2007 Showed that using action-based distance results in plans that are likely to be also diverse with respect to behavior and causal structure LPG can scale-up well to large problems with the proposed changes The approach and results are representative of how other planners may be modified to find diverse plans Domain Independent Approaches for Finding Diverse Plans 20 Future Work On the same thread Generalized problem Jan 09, 2007 Solution approaches for more problems Extensive experiments More suitable distance measures Other action representations: Nondeterministic, HTN actions, … Plans with different goals Domain Independent Approaches for Finding Diverse Plans 21 Appendix Jan 09, 2007 Domain Independent Approaches for Finding Diverse Plans 22 Purpose for Comparison and Characteristics of the Plan Distance Measure Plans for visualization purpose Plans for execution purpose Jan 09, 2007 Minimal and non-minimal plans should be found similar. They achieve the goal, after all! Plans for different goals should be seen different Minimal and non-minimal plans should be found different. Plans with similar execution trace should be seen similar even if they are for different goals Domain Independent Approaches for Finding Diverse Plans 23