1. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati 10/4: Knowledge-based Planning  Semester Project Proposals due

2. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Life must be lived forward, but can only be understood backward.--Kierkegaard

3. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati CUSTOMIZING PLANNERS WITH DOMAIN SPECIFIC KNOWLEDGE 1.User-assisted customization “Planning by Cheating”?? (accept domain-specific knowledge as input) 2. Automated customization (learn regularities of the domain through analysis or experience)

4. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Digression: Domain-Independent vs. Domain Specific vs. Domain Customized G Domain independent planners only expect as input the description of the actions (in terms of their preconditions and effects), and the description of the goals to be achieved G Domain dependent planners make use of additional knowledge beyond action and goal specification –Domain dependent planners may either be stand alone programs written specifically for that domain OR domain independent planners customized to a specific domain –In the case of domain-customized planners, the additional knowledge they exploit can come in many varieties (declarative control rules or procedural directives on which search choices to try and in what order) –The additional knowledge can either be input manually or in some cases, be learned automatically Unless noted otherwise, we will be talking about domain-independent planning Review

5. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Improving Performance through Customization G Biasing the search with control knowledge acquired from experts –Non-primitive actions and reduction schemas –Automated synthesis of customized planners »Combine formal theory of refinement planning and domain-specific control knowledge G Use of learning techniques –Search control rule learning –Plan reuse

6. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati User-Assisted Customization (using domain-specific Knowledge) G Domain independent planners tend to miss the regularities in the domain G Domain specific planners have to be built from scratch for every domain An “Any-Expertise” Solution: Try adding domain specific control knowledge to the domain- independent planners ACME all purpose planner Ronco Blocks world Planner Ronco logistics Planner Ronco jobshop Planner AC-RO Customizable planner Domain Specific Knowledge

7. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Task Decomposition (HTN) Planning G The OLDEST approach for providing domain-specific knowledge –Most of the fielded applications use HTN planning G Domain model contains non-primitive actions, and schemas for reducing them G Reduction schemas are given by the designer –Can be seen as encoding user-intent »Popularity of HTN approaches a testament of ease with which these schemas are available? G Two notions of completeness: –Schema completeness » (Partial Hierarchicalization) –Planner completeness

8. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Modeling Reduction Schemas GobyBus(S,D) t1: Getin(B,S) t2: BuyTickt(B) t3: Getout(B,D) In(B) Hv-Tkt Hv-MoneyAt(D)

9. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati A concretization is a task network produced by repeatedly reducing a non-primitive task until all tasks are primitive

10. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Modeling Action Reduction Affinity between reduction schemas and plan-space planning

11. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Dual views of HTN planning G Capturing hierarchical structure of the domain –Motivates top-down planning »Start with abstract plans, and reduce them G Many technical headaches –Respecting user-intent, maintaining systematicity and minimality [Kambhampati et. al. AAAI-98] »Phantomization, filters, promiscuity, downward- unlinearizability.. G Capturing expert advice about desirable solutions –Motivates bottom-up planning »Ensure that each partial plan being considered is “legal” with respect to the reduction schemas »Directly usable with disjunctive planning approaches G Connection to efficiency is not obvious Relative advantages are still unclear... [Barrett, 97]

12. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati

13. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati

14. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati HTN vs. T-STN G Nau’s book talks about a “special” version of HTN called T-STN (Total- ordering STN) G T-STN schemas are HTN schemas where all the ordering relations between tasks are constrained to be “contiguity” constraints –This means that the concretizations of different tasks cannot be interleaved –Bug or feature? »From the planner’s point of view, it can be a feature Most of the difficulties in HTN planning come from the need to do plan-space refinement in the presence of non-primitive tasks (phantomization, conflict resolution etc..); eliminating that will make HTN planning a no- brainer »BUT from the domain-writer’s point of view it is clearly a BUG l Would have to write many more schemas –Same is true of Yang’s unique main sub-action thing (although it is less drastic than T-STN) G The only HTN planners that use STN schemas are those from Nau’s group … –SIPE and O-Plan use schemas with precedence orderings

15. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati

16. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati

17. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati HTN vs. STRIPS planning: Competing or Complementary? G Much of the discussion till now has been done with the assumption that HTN is just a way of doing STRIPS planning G HTN can also be seen as solving a different problem –Can develop plans at various levels of abstraction –Can model domains with partial domain knowledge »What you consider “primitive” action may be changeable l What you thought was a correct plan may not be..

18. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati 9/6

19. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati

20. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati

21. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Changes to Plan-Space Refinement (in the presence of non-primitive tasks)

22. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati

23. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Full procedural control: The SHOP way Travel by bus only if going by taxi doesn’t work out Shop provides a “high-level” programming language in which the user can code his/her domain specific planner -- Similarities to HTN planning Uses T-STN schemas -- Order of use of reduction schemas can be specified The SHOP engine can be seen as an interpreter for this language [Nau et. al., 99] Blurs the domain-specific/domain-independent divide How often does one have this level of knowledge about a domain?

24. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Non-HTN Declarative Guidance Invariants: A truck is at only one location at(truck, loc1, I) & loc1 != loc2 => ~at(truck, loc2, I) Optimality: Do not return a package to a location at(pkg, loc, I) & ~at(pkg,loc,I+1) & I ~at(pkg,loc,j) Simplifying: Once a truck is loaded, it should immediately move ~in(pkg,truck,I) & in(pkg,truck,I+1) & at(truck, loc, I+1) => ~at(truck, loc, I+2) Once again, additional clauses first increase the encoding size but make them easier to solve after simplification (unit-propagation etc). [Kautz & Selman, AIPS-98]

25. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Case Study: SATPLAN with domain specific knowledge G Instantiate the domain specific rules as well, and then add them to the encoding. G Solve the full encoding… –Won’t the encoding size increase with domain spefic rules? »Yes, but the new rules support a lot of simplification so that after simplification, the encoding size actually reduces!! Example in Blocks world (from Kautz & Selman, AIPS-99) BW-c: Original: 4258 vars (10% increase) / 30986 clauses (300% increase) Final (after simplification): 3526 vars 22535 clauses

26. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati SAT encodings of HTN planning G Abstract actions can be seen as disjunctive constraints –K-step encoding has each of the steps mapped to a disjunction of the non-primitive tasks –If a step s is mapped to a task N, then one of the reductions of N must hold (**The heart of encoding setup**) –+ The normal constraints of primitive action-based encoding »Causal encodings seem to be a natural fit (given the causal dependencies encoded in reduction schemas) [Mali & Kambampati, AIPS-98] Constraints from action-based encodings HTN constraints HTN-compliant Solutions Solutions for action based encodings

27. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Solving HTN Encodings [Kambhampati & Mali, AIPS-98] # Clauses TIME Puzzle: How can increasing encoding sizes lead to efficient planning? Abstract actions and their reductions put restrictions on the amount of step-action disjunction at the primitive level. --Reduction in step-action disjunction propagates e.g. Fewer causal-link variables, Fewer exclusion clauses… Savings won’t hold if each non-primitive task has MANY reductions

28. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Rules on desirable State Sequences: TLPlan approach TLPlan [Bacchus & Kabanza, 95/98] controls a forward state-space planner Rules are written on state sequences using the linear temporal logic (LTL) LTL is an extension of prop logic with temporal modalities U until [] always O next <> eventually Example: If you achieve on(B,A), then preserve it until On(C,B) is achieved: [] ( on(B,A) => on(B,A) U on(C,B) )

29. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Keep growing “good” towers, and avoid “bad” towers Good towers are those that do not violate any goal conditions TLPLAN Rules can get quite boroque How “Obvious” are these rules? The heart of TLPlan is the ability to incrementally and effectively evaluate the truth of LTL formulas.

30. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Control rules for choice points G UCPOP developed a language in which user can write control rules that tell the planner what to do when it faces a specific sort of choice (e.g. between open conditions; establishers for open conditions etc.). G HSTS—a metric temporal planner that NASA used in its DeepSpace mission used similar control rule language. –The user needs to know the innards of the planner to write these rules –The rules are hard to debug »Learning such rules automatically is a way out

31. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Folding the Control Knowledge into the planner: CLAY approach G Control knowledge similar to TLPlan’s G Knowledge is folded using KIDS semi-automated software synthesis tool into a generic refinement planning template –Use of program optimizations such as »Finite differencing »Context-dependent & independent simplification G Empirical results demonstrate that folding can be better than interpreting rules Srivastava & Kambhampati, JAIR 97 Caveat: Current automated software synthesis tools have a very steep learning curve

32. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Many User-Customizable Planners G Conjunctive planners –HTN planners »SIPE [Wilkins, 85-] » NONLIN/O-Plan [Tate et. al., 77-] »NOAH [Sacerdoti, 75] »Also SHOP (Nau et. al., IJCAI-99) –State-space planners »TLPlan [Bacchus & Kabanza, 95; 99] »TALPlan [Kvarnstrom & Doherty, 2000] –Customization frameworks »CLAY [Srivastava & Kambhampati, 97] G Disjunctive planners –HTN SAT [Mali & Kambhampati, 98] –SATPLAN+Dom [Kautz & Selman, 98] How do they relate? What are the tradeoffs?

33. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati With right domain knowledge any level of performance can be achieved... G HTN-SAT, SATPLAN+DOM beat SATPLAN… –Expect reduction schemas, declarative knowledge about inoptimal plans G TLPLAN beats SATPLAN, GRAPHPLAN –But uses quite detailed domain knowledge G SHOP beats TLPLAN…(but not TALPlan) –Expects user to write a “program” for the domain in its language »Explicit instructions on the order in which schemas are considered and concatenated Who gets the credit? -The provider of domain knowledge? -The planner that is capable of using it?

34. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Types of Guidance G Declarative knowledge about desirable or undesirable solutions and partial solutions (SATPLAN+DOM; Cutting Planes) G Declarative knowledge about desirable/undesirable search paths (TLPlan & TALPlan) G A declarative grammar of desirable solutions (HTN) G Procedural knowledge about how the search for the solution should be organized (SHOP) G Search control rules for guiding choice points in the planner’s search (NASA RAX) Planner specific. Expert needs to understand the specific details of the planner’s search space (largely) independent of the details of the specific planner [affinities do exist between specific types of guidance and planners)

35. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Where does guidance come from? G Learned from (planning) experience –EBL-based search control rule learning »(UCPOP+EBL), PRODIGY –inductive learning –Case-based reasoning (DerSNLP) G Given by humans (often, they are quite willing!) –As declarative rules (HTN Schemas, Tlplan rules) »Don’t need to know how the planner works.. »Tend to be hard rules rather than soft preferences… »Whether or not a specific form of knowledge can be exploited by a planner depends on the type of knowledge and the type of planner –As procedures (SHOP) »Direct the planner’s search alternative by alternative.. how easy is it to write control information?

36. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Ways of using the Domain Knowledge G As search control –HTN schemas, TLPlan rules, SHOP procedures –Issues of Efficient Matching G To prune unpromising partial solutions – HTN schemas, TLPlan rules, SHOP procedures – Issues of maintaining multiple parses G As declartative axioms that are used along with other knowledge –SATPlan+Domain specific knowledge –Cutting Planes (for ILP encodings) –Issues of domain-knowledge driven simplification G Folded into the domain-independent algorithm to generate a new domain-customized planner –CLAY –Issues of Program synthesis

37. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Conundrums of user-assisted cutomization G Which planners are easier to control? –Conjunctive planners are better if you have search control knowledge »Forward State Space (according to TLPlan) »Plan-space planners (according to HTN approaches) – Disjunctive planners are better if your knowledge can be posed as additional constraints on the valid plans »Which SAT encoding? l HTN knowledge is easier to add on top of causal encodings G Which approach provides the best language for expressing domain knowledge for the lay user? –(Mine--no, no, Mine!) G What type of domain knowledge is easier to validate? G When does it become “cheating”/ “wishful-thinking” –Foolish not to be able to use available knowledge –Wishful to expect deep procedural knowledge...

38. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati “Evaluating” KB-planners G Since the 2 nd IPC, there has been a separate track for “knowledge-based planners” at the competition. –Basically only the TLPlan (and its variant TALPlan) and SHOP took part (in comparison close to a dozen planners take part in the domain-independent track) »In 2000, TALPlan won »In 2002, TLPlan won –Quite controversial G When SHOP does better than TLplan on Gizmo domain, what does this mean? –That SHOP is a better planning algorithm than TLplan? –That Dana Nau knows Gizmo domain better than Faheim Bacchus? –That the easily available control knowledge in Gizmo domain is more naturally representable as TLplan rules than SHOP schemas?

39. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Are we comparing Dana & Fahiem or SHOP and TLPlan? (A Critique of Knowledge-based Planning Track at ICP) Subbarao Kambhampati Dept. of Computer Science & Engg. Arizona State University Tempe AZ 85287-5406

40. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati The “I am not an anti-dentite” Disclaimers.. G I think KB planning is a swell idea –I started my career with HTN planning… G I think the KB planning track at IPC is a swell idea –has done more to increase interest in KBplanning than the bi-annual polemics and laments about lack of interest in “Knowledge-based planning” G I think Fahiem and Dana are REALLY swell –(in case they don’t buy that) I may already have a black-belt in Karate.. I’d rather learn from one bird how to sing than teach ten thousand stars how not to dance --e.e. cummings

41. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati What are the lessons of KB Track? G If TLPlan did better than SHOP in ICP, then how are we supposed to interpret it? –That TLPlan is a superior planning technology over SHOP? –That the naturally available domain knowledge in the competition domains is easier to encode as linear temporal logic statements on state sequences than as procedures in the SHOP language? –That Fahiem Bacchus and Jonas Kvarnstrom are way better at coming up with domain knowledge for blocks world (and other competition domains) than Dana Nau? We are NOT asking the right questions

42. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Questions worth asking in KB planner comparisons (IMHO) G How easy/natural (for humans) is the language in which the planner accepts control knowledge? G How easy is it to “validate” the control knowledge being input to the planner? G Is the naturally available knowledge about a specific domain easily encoded in the language accepted by the planner? G Does the planner allow “any expertise” behavior— solving the problems even without any control knowledge, but improving performance with added control knowledge? (or is the control knowledge tightly intertwined with the domain physics?).

43. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati How/Why the competition is not asking the right questions… G The role of the knowledge-engineer is played by the same person(s) who wrote the planner. So, the question of how natural the specific language is for third-party knowledge engineers is largely unaddressed. G No reasonable time limits are placed on coming up with the control knowledge. So, we don’t learn much (or anything) about whether or not naturally available knowledge about a domain is easily representable in the language accepted by the planner.

44. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Some suggestions for change… G Recruit third-party volunteers who will play the role of knowledge engineers for the KB planners. –Ideally, we would like to have the same people writing the control knowledge for a given domain for all the competing approaches (so one knowledge engineer per domain rather than one knowledge engineer per planner). G (Alternative to above) Specify the control knowledge that is available, so all planners encode the same general knowledge. –One idea might be to ask the designers of the domains (e.g. David Smith and his cohorts for the Satellite and Rovers domain) to provide, in english, what sort of control information they would like the planner to use. G Measure the time taken to write and validate the control knowledge. G Analyze the knowledge encoded by the different KB planners for the same domain –Characterize it in terms of (a) whether the knowledge is procedural or declarative and (b) how hard would it be to “learn” the same knowledge.

45. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Slides from here-on about “learning” domain knowledge were not used

46. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Automated Customization of Planners G Domain pre-processing – Invariant detection; Relevance detection; Choice elimination, Type analysis »STAN/TIM, DISCOPLAN etc. »RIFO; ONLP G Abstraction »ALPINE; ABSTRIPS, STAN/TIM etc. G Learning Search Control rules »UCPOP+EBL, »PRODIGY+EBL, (Graphplan+EBL) G Case-based planning (plan reuse) »DerSNLP, Prodigy/Analogy Most of the work has been done in the context of Conjunctive Refinement Planners Read Terry Zimmerman’s “Learning assisted planning: Looking back, Taking Stock and Going Forward”

47. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Symmetry & Invariant Detection G Generate potential invariants and test them –DISCOPLAN [Gerevini et. al.] »Allows detection of higher-order mutexes –Rintanen’s planner »Uses model-verification techniques –STAN/TIM »Type analysis of the domain is used to generate invariants –ONLP (Peot & Smith) »Use operator graph analysis to eliminate non-viable choices

48. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Abstraction G Idea –Abstract some details of the problem or actions. –Solve the abstracted version. –Extend the solution to the detailed version G Precondition Abstraction –Work on satisfying important preconditions first »Importance judged by: l Length of plans for subgoals [ABSTRIPS, PABLO] l Inter-goal relations [ALPINE] l Distribution-based [HighPoint] –Strong abstractions (with downward refinement property) are rare –Effectiveness is planner-dependent »Clashes with other heuristics such as “most constrained first”

49. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Example: Abstracting Resources G Most planners thrash by addressing planning and scheduling considerations together –Eg. Blocks world, with multiple robot hands G Idea: Abstract resources away during planning –Plan assuming infinite resources –Do a post-planning resource allocation phase –Re-plan if needed (with Biplav Srivastava)

50. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Learning Search Control Rules with EBL UCPOP+EBL Explain leaf level failures Regress the explanations to compute interior node failure explanations Use failure explanations to set up control rules Problems: -- Most branches end in depth-limits >No analytical explanation >Use preference rules? -- THE Utility problem >Learn general rules >Keep usage statistics & prune useless rules If Polished(x)@S & ~Initially-True(Polished(x)) Then REJECT Stepadd(Roll(x),Cylindrical(x)@s) (Kambhampati, Katukam, Qu, 95)

51. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Example Rules (Learned) If Polished(x)@S & ~Initially-True(Polished(x)) Then REJECT Stepadd(Roll(x),Cylindrical(x)@s) UCPOPProdigy If (and (current-node node) (candidate-goal node (shape obj Cyl)) (candidate-goal node (surface-condition obj polished))) Then: Prefer (shape obj cyl) to (surface-condition obj polished) Pruning rule Preference rule

52. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Case-based Planning Macrops, Reuse, Replay G Structures being reused –Opaque vs. Modifiable –Solution vs. Solving process (derivation) G Acquisition of structures to be reused –Human given vs. Automatically acquired G Mechanics of reuse –Phased vs. simultaneous G Costs –Storage & Retrieval costs; Solution quality

53. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Case-study: DerSNLP G Modifiable derivational traces are reused G Traces are automatically acquired during problem solving –Analyze the interactions among the parts of a plan, and store plans for non-interacting subgoals separately »Reduces retrieval cost –Use of EBL failure analysis to detect interactions G All relevant trace fragments are retrieved and replayed before the control is given to from-scratch planner –Extension failures are traced to individual replayed traces, and their storage indices are modified appropriately »Improves retrieval accuracy ( Ihrig & Kambhampati, JAIR 97) Old cases

54. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati DerSNLP: Results Performance with increased Training % Solvability with increased traning (JAIR, 97) Library Size 5 3 1 Problem size (#goals)

55. MA2: Recent Advances in AI Planning: A Unified View Subbarao Kambhampati Reuse in Disjunctive Planning G Harder to make a disjunctive planner commit to extending a specific plan first G Options: –Support opaque macros along with primitive actions »Increases the size of k-step disjunctive plan »But a solution may be found at smaller k –Modify the problem/domain specification so the old plan’s constraints will be respected in any solution (Bailotti et. al.) –MAX-SAT formulations of reuse problem »Constrain the encoding so that certain steps may have smaller step-action mapping and ordering choices »Causal encodings provide better support [with Amol Mali]