Synthesizing Robust Plans under Incomplete Domain Models Tuan Nguyen and Subbarao Kambhampati Department of CSE, Arizona State University Minh Do Palo Alto Research Center Acknowledgement: Funding from ONR grants N00014-09-1-0017, N00014-07-1-1049, NFS grant IIS-0905672, and by DARPA and the U.S. Army Research Laboratory under contract W911NF-11-C-0037. Planning – The Traditional View Problem instance a1 a2 PLANNER a3 Domain model “Valid” plan Deterministic actions Stochastic non-deterministic actions add “p” delete “q” “Probabilistic” plan/policy 0.4 : add “p” 0.6 : delete “q” COMPLETE/FULL MODEL Laborious and error-prone!!! Model-lite Planning as Generalized Planning Model-lite Planning A Domain Incompleteness View Deterministic actions Stochastic, non-deterministic actions I/O types Task dependency (e.g. workflows management, web service composition) Missing some preconditions/effects of actions (e.g. Garland & Lesh, AAAI-05) There are known knowns; there are things we know that we know. There are known unknowns; that is to say, there are things that we now know we don’t know. But there are also unknown unknowns; there are things we do not know we don’t know. Problem Formulation Incomplete domain ̃D = 〈 F , A〉 Proposition setF = { p1 , p 2 ,... , p n } Actiona∈ A Add (a) q p Pre(a) a ̃Pre(a) p ' w pre a ( p') - - Del (a) r q' ̃ Add (a) w del a (q ' ) r' del a w (r ' ) ̃Del (a) Different from stochastic effects! Problem Formulation Transition function π I ̃D ? ̃D Completion set 〈〈 ̃D 〉〉 D1 D2 D2 K π I Di π γ (π, I , ̃D )= γ (π, I , Di ) U D ∈〈〈 ̃D 〉〉 i Dj Assumption: “Inapplicable” action causes state unchanged Challenges Language for domain incompleteness A robustness measure for plans Generating robust plans Incompleteness Annotation: Modeling Issues Incompleteness annotations can be at Schema level Grounded level Or in between Incompleteness Annotation: Modeling A tourist planning to Issues have food in a small Restriction on variable values q( x1 ) p( x 3) a( x 1 , x 2 , x 3 ) - r ( x2) town is not sure if she needs to have cash. Her action have_food(M: Meals, C: Town) has possible precondition need_cash(M: Meals, C:Town): when (C=the town) q ' ( x1 ,C 3 ) p ' (C 1 , x 2 ) - Possible add: q ' ( x1 , x3 ): when( x 3 = C 3 ) Possible precondition: p ' ( x1 , x 2 ): when ( x 1= C 1 ) The domain writer knows that p ' ( x1 , x 2 ) is NOT a precondition of a( x 1 , x 2 , x 3 ) when x1≠ C 1 , and may be in other cases r ' ( x 1 ,C 2 ) Possible delete: r ' ( x 1 , x 2): when( x 2= C 2 ) Incompleteness Annotation: Modeling Issues Restriction on variables: Possible preconditions/effects depending on values of some variables, but such values are have_food(M: Meals, q( x1 ) unknown! C: Town) has possible p( x 3) p ' ( x 1 , x 2) a( x 1 , x 2 , x 3 ) Possible precondition: p ' ( x1 , x 2 ): depends x1 p ' ( x1writer , x 2 ):knows depends The domain that xp1' ( x1 , x 2 ) is a possible precondition of a( x 1 , x 2 , x 3 ) when x1 has some specific value, but unknown. - r ( x2) q ' ( x1 , x 3 ) - precondition need_cash(M: Meals, C:Town): depends C Possible add: q ' ( x1 , x3 ): depends x 3 r ' ( x 1 , x 2) Possible delete: r ' ( x 1 , x 2): depends x 2 Incompleteness Annotation: Modeling Issues Correlated incompleteness p Pre(a) a ̃Pre(a) p ' w pre a ( p') - - q Add (a) r Del (a) q' ̃ Add (a) w del a (q ' ) r' del a w (r ' ) If p' is realized as a precondition of a, then more likely that r' will be delete effect of the action. ̃Del (a) Challenges Language for domain incompleteness A robustness measure for plans Generating robust plans A Robustness Measure for Plans A plan in ̃D may fail or succeed in reaching ̃D goal states Completion set 〈〈 ̃D 〉〉 Plan execution reaches goal state? D1 D2 D2 yes no no Plan robustness: Cummulative probability mass of complete models under which the plan succeeds. K A Spectrum of Robust Planning Problems Robustness assessment Maximally robust plan generation Generating plans with desired level of robustness Cost sensitive robust plan generation Incremental robustification Challenges Language for domain incompleteness A robustness measure for plans Generating robust plans to Conformant Probabilistic Planning Problem Conformant Probabilistic Planning Problem Domain D' = 〈 F ' , A' 〉 Proposition set F ' Action a ' ∈ A' Mutually exclusive and exhaustive Preconditions Pre(a ' )⊆ F ' Conditional effects e= (cons(e) , O(e)= {( Pr (ϵ) , add (ϵ) , del (ϵ))}) 0.7 0.3 a' 0.2 0.8 Problem P ' = 〈 D ' ,b I ,G ' ,ρ〉 Compilation Approach: An Example Compiled “pick-up” Compilation Example Correctness of the compilation Experimental Results Logistics Two cities C 1 and C 2 each with a downtown and airport. Heavy packages at the downtown areas Robots Ri ,1 , ... , Ri , m at the airport of the city C i Source of incompleteness: robots R1, j , R 2, j were made from the same manufacturer, having possible precondition that packages should not be heavy to pick. Goals: move packages from C 1 to C 2 and vice versa. Plans can be made more robust by using robots from different manufacturers after moving them into the downtown area, with the cost of increasing the plan length. Experimental Results Satellite Two satellites S 1 and S 2 orbiting the planet Earth Imagers Li ,1 ,... , Li , m installed on S i Source of incompleteness: lense of L1, j , L 2, j were made from the type of material M j and can produce possible effect that images taken are mangled. Goals: images taken in some mode at some direction. Plans can be made more robust by using additional instruments, which might be in different satellites, but should be of different types of material and can also take an image of the interested mode at some direction. Experimental Results Logistics Satellite Number of m models: 2 Observations Fixed the number of models: Plan tends to be longer with increasing robustness threshold Fixed the robustness threshold: The maximal robustness value of plans that can be returned increases with higher number of manufacturers. Related work K-faults plans (Jensen et al 2004) Plan evaluation with incomplete models (Garland & Lesh, 2002) Planning and Acting in Incomplete Domains (Bryce & Weber, 2011) Robust temporal planning (Fox 2006) Handling incompleteness at the atomic level: MDP with uncertain transition probabilities (Satia & Lave 1973; Delago & Sanner 2009) Bounded parameter MDP (Givan, Leach, Dean 2000) Conclusion & Future work Introduce planning with incomplete models Incompleteness annotations Robustness measure for plans A spectrum of robust planning problems Finding a plan with at least a robustness value: compilation approach Future work: Heuristic approach utilizing annotations Plan robustification, and other problems in the spectrum. Thank you! Q&A Backup Uniform distribution: 6/8 robustness 0.7 0.3 Belief state b a' 0.2 0.8 Resulting belief state b' R(π, ̃P )= ∑ D i ∈ 〈〈 ̃D 〉〉 , γ (π, I , G)∣%equal G Pr ( Di ) Problem Formulation q Pre(a) p a p' - r q' r' Problem Formulation q Add (a) r Del (a) q' ̃ Add (a) p Pre(a) a ̃Pre(a) p ' pre wa ( p ' ) - - del wa ( p ' ) r' del a w ( p' ) ̃Del (a)