1. Haley: A Hierarchical Framework for Logical Composition of Web Services Haibo Zhao, Prashant Doshi LSDIS Lab, Dept. of Computer Science, University of Georgia IEEE International Conference on Web Services 2007

2. Outline  Introduction  Motivating scenario  Background  Model: First order Semi-Markov decision processes (FO-SMDP)  Composing nested Web processes using Haley  Architecture  Experiment & Discussion

3. Introduction  Web service composition Business processes with Web services as components  Existing approaches to composition: AI planning Classical planning techniques  Golog, Model checking-based planning, HTN planning, Synthy Decision-theoretic planning  MDP (Doshi2004)  Limitations Classical planning assumes deterministic behavior of Web services Guarantee correctness but not optimality State space explosion Cannot operate directly on WS descriptions in FOL

4. Our Approach: Haley  Stochastic SMDP model Handle uncertainties in WS invocation Provide cost-based optimality  Hierarchical model to represent the hierarchies in Web processes Address the scalability problem  FOL based representation Directly operate on WS description in FOL

5. Handling Orders in Supply Chain Level 1: Composition using composite FO-SMDP Abstract action Level 0: Composition using primitive FO-SMDP Level 0: Composition using primitive FO-SMDP Abstract action

6. Background: Probabilistic Situation Calculus [Reiter01]  A FOL based framework for representing actions, changes and reasoning about them  Probabilistic Situation calculus elements Actions: parameterirzed FO terms ReceiveOrder(o) Situations: sequence of actions representing the state of the world do(ReceiveOrder(o),s 0 ) Fluents: situation-dependent relations and functions whose truth values may vary HaveOrder(o, s) Nature’s Choices: capture stochastic results of actions Choice(CheckCustomer(o), a) ≡ a = CheckCustomerS(o) ∨ a = CheckCustomerF(o) Probabalities for nature’s choices: Pr(CheckCustomerS(o),CheckCustomer(o), s) = 0.9 Pr(CheckCustomerF(o),CheckCustomer(o), s) = 0.1 Precondition Axioms: HaveOrder(o, s) ⇒ Poss(CheckCustomer(o), s) Successor state Axioms: Describe the effects on fluents Poss(a, s) ⇒ HaveOrder(o, do(a, s)) ⇔ a = ReceiveOrderS(o) ∨ (HaveOrder(a, s) ∧ a ≠CancelOrderS(o))

7. First Order MDP(FO-MDP) [Boutilier 01]  Probabilistic situation calculus representation allows concise specification of complex domains  Specify lump sum reward/cost and utilities with case notation kCase(A(x)) = case[ A(x) = CheckCustomer(o), 2; A(x)= VerifyPayment(o), 3; A(x) = ChargeMoney(o),2 ]  Avoid explicit state and action enumeration  A decision-theoretic regression algorithm for solving FO-MDPs

8. First Order Semi-MDPs (FO-SMDP)  FO-SMDP is a temporal generalization of FO-MDPs: The sojourn time of actions are modeled with a density function; and the system will incur an action-duration cost at an accumulating rate  Case notation of sojourn time distribution  Case notation of accumulating rate  Total reward of a state-action pair  Representing total reward and utilities with case notation, FO-SMDP can be solved analogously to FO-MDP using DT regression

9. Level 1: Composition using composite FO-SMDP Abstract action Level 0: Composition using primitive FO-SMDP Level 0: Composition using primitive FO-SMDP

10. Elicitation of Model Parameters (level 0) Level 0: Model parameters may be obtained from WSDL-S/SAWSDL, OWL-S descriptions of Web services, and service level agreements  Compile situation calculus axioms from preconditions and effects e.g. WS: ChargeMoney(o) Precondition: V alidCustomer(o) AND V alidPayment(o) Effect: Charged(o) The precondition axiom: ValidCustomer(o, s) ∧ V alidPayment(o, s) ⇒ Poss(ChargeMoney(o), s) The successor state axiom: Poss(a, s) ⇒ Charged(o, do(a, s)) ⇔ a=ChargeMoneyS(o) ∨ Charged(o, s)  Elicit non-functional parameters from service level agreement:

11. Deriving Model Parameters for Abstract Actions (level≥1) Level ≥1: Derive model parameters related to abstract actions from lower level Web process  We need to know successor state axioms and the case notations of lump sum cost, sojourn time distribution and accumulating rate

12. Deriving Model Parameters for Abstract Actions  Successor state axioms –Let –We have And –The successor state axiom of VerifyOrder(o) becomes: Relation between high-level fluents and low-level fluents Relation between high-level abstract actions and low-level actions

13. Deriving Model Parameters for Abstract Actions  Lump sum cost K –lump sum cost of the abstract action is the total of lump sum costs of the corresponding primitive actions –Add a new case into the case notation of K k VO = kCase(CheckCustomer(o)) + kCase(VerifyPayment(o)) + kCase(ChargeMoney(o)) New case to be added

14. Deriving Model Parameters for Abstract Actions  Sojourn time distribution F –Assume the sojourn time of all primitive actions follows Gaussian distribution: f CC (t)=N(t; µ cc, σ cc ), f vo (t)=N(t; µ vo, σ vo ) and f cm (t)=N(t; µ cm, σ cm ) –Linear combination of Gaussian distributions is a Gaussian distribution, the abstract action VerifyOrder also follows Gaussian f vo (t)=N(t; µ vo, σ vo ) where: –Add a new case into the case notation of F New case to be added

15. Deriving Model Parameters for Abstract Actions  Cost Accumulating Rate C –Accumulated cost of an abstract action is the total accumulated cost of all corresponding primitive actions –Add a new case into case notation of C Given model parameters for abstract actions, composite FO-SMDP can be solved analogously to a primitive FO-SMDP

16. Architecture of Haley

17. Interleaved Generation and Execution of Nested Web Process

18. Performance Evaluation − Comparison with HTN planning and MBP planning on supply chain scenario − Execute the processes generated by three approaches in a simulated environment 1000 times, measure average rewards The performance of HTN approaches ours as the environment becomes less uncertain; Haley provides cost based optimization compared to MBP planner

19. Performance Evaluation Comparisons of different decision theoretic planners in the same domain and the collected runtimes Hierarchal decomposition significantly improves the performance First order representation avoids the explicit state enumeration

20. Discussion Many AI planning based approaches  AI classical planning is not designed to handle WS composition  Assumes deterministic behavior of Web services  Cannot directly operate on WS descriptions in FOL  Does not scale well to large problems Haley: our hierarchical framework  Stochastic optimization manages uncertainty and delivers optimality  Able to operate directly on WS descriptions in FOL  Exploits hierarchy  scalability  Better performance in uncertain environments Future work  Incorporate data mediation in Haley

21. Thank You! Questions? Contact us Haibo Zhao: zhao@cs.uga.edu Prashant Doshi: pdoshi@cs.uga.edu