220 likes | 239 Views
4 th International Conference on Service Oriented Computing. A Hierarchical Framework for Composing Nested Web Processes. Haibo Zhao, Prashant Doshi LSDIS Lab, Dept. of Computer Science, University of Georgia. Introduction. Web process Business processes with Web services as components
E N D
4th International Conference on Service Oriented Computing A Hierarchical Framework for Composing Nested Web Processes Haibo Zhao, Prashant Doshi LSDIS Lab, Dept. of Computer Science, University of Georgia
Introduction • Web process • Business processes with Web services as components • Existing approaches to Composition: AI planning • Classical planning techniques • Golog (McIlraith2001), MBP-based planning (Traverso2003), HTN planning (Wu2003), Synthy (Srivastava2005) • Decision-theoretic planning • MDP (Doshi2004) • Limitations • Classical planning assumes deterministic behavior of Web services • Guarantees correctness but not optimality • Existing approaches do not scale well to large processes • Our approach • Hierarchy in Web processes to address the scalability problem • A hierarchical semi-MDP to model the hierarchy • Stochastic SMDP model to handle uncertainties and optimality
Motivating Scenario 1Order handling in supply chain singly nested
Motivating Scenario 2 Patient Transfer Pathway doubly nested
Our Approach Level 0 Composition using primitive SMDP Level 1 Composition using composite SMDP Abstract actions Level 0 Composition using primitive SMDP
Semi-MDP (SMDP) SMDP = <S, A, T, K, F, C, s0> • S: set of all states Capture feature-based state space (factored into variables) E.g. isOrderValid Yes/No/Unknown • A: set of actions, some temporally extended Model Web service invocations or operations • Level 0 – action is the invocation of a primitive WS E.g.receiveOrder • Level > 1 – Abstract action is the invocation of a lower-level Web process E.g. verifyOrder • T: transition function, T: S X A Δ(S) Uncertain effects with probabilitiesE.g. T( inventoryAvail = Yes | Check Inventory Status , inventoryAvail = Unknown) = 0.3 • K: lump sum reward/cost, K: S X A R WS invocation or usage reward/cost • F: sojourn time distribution, F: S X A Δ(t) • Level 0 – uncertain response time of Web services • Level > 1 – uncertain runtime of lower level Web process • C: reward/cost accumulating rate, C: S X A RReward/cost rate (cost per time unit) of using Web services • s0 : initial state
Solving SMDPs • Expected Utility of state s: Quantitative measure of immediate effects and long-term effects of actions Where: • Solving a SMDP is maximizing the expected utility of each state, Solution techniques include: Value Iteration, Policy Iteration and Linear Programming • The solution is a policy: Mapping from states of process environment to actions • More robust than a sequence of Web service invocations • Execution of a policy: 1. Determine the current state s 2. Invoke WS given by action a based on the policy 3. Repeat 1-2 until the goal is achieved
Elicitation of Model Parameters • Level 0: Model parameters may be obtained from WSDL\SAWSDL, OWL-S descriptions of Web services, and WS-Agreements • Level >1: Derive model parameters related to abstract actions from lower level Web process • Specifically, we want to know transition function T, lump sum reward K, sojourn time distribution F, and accumulating rate C
Abstract Actions and Variable Correspondence • Abstract action “Verify Order”( ) is composed of “Check Customer”(acc), “Verify Payment”(avp), and “Charge Money”(acm) • Actions affect only certain variables • Correspondence between high-level and lower-level preconditions and effects • For example:
Deriving Transition Function T for Abstract Actions • Correspondence: E.g. The transition from (OV=U) to (OV=Y)
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
Deriving Model Parameters for Abstract Actions • Sojourn time distribution F • Assume the sojourn time of all primitive actions follows Gaussian distribution: FCC: N(t; µcc, σcc), Fvo: N(t; µvo, σvo) and Fcm: N(t; µcm, σcm) • Linear combination of Gaussian distributions is a Gaussian distribution • The abstract action VerifyOrder also follows Gaussian Fvo: N (t; µvo, σvo) where:
Deriving Model Parameters for Abstract Actions • Accumulating Cost Rate C • Accumulated cost of an abstract action is the total accumulated cost of all corresponding primitive actions • Where: Ecc(F), Evp(F) and Ecm(F) are expected sojourn time Given model parameters for abstract actions, composite SMDP can be solved analogous to a primitive SMDP
Performance Evaluation • Methodology • Comparison with HTNs (Wu 2003) on two scenarios • Run the processes generated by two approaches in a simulated environment 1000 times • Measure average reward and standard deviation The performance of HTN approaches ours as the environment becomes less uncertain
Discussion • Many AI planning approaches • AI classical planning is not designed to handle WS composition • Assumes deterministic behavior of Web services • Does not scale well to large problems • Our hierarchical framework • Stochastic optimization manages uncertainty and delivers optimality • Exploits hierarchy scalability • Better performance in uncertain environments • Future work • Integrate first-order logic to manage state space explosion
Outline • Introduction • Motivating scenarios • Semi-Markov decision process (SMDP) • Composing nested web processes using Hierarchical SMDP • System architecture • Experiment & Discussion