Optimal Adaptation in Web Processes with Coordination Constraints

1
Optimal Adaptation in Web Processes with
Coordination Constraints

• Kunal Verma, Prashant Doshi, Karthik Gomadam,
  John A. Miller, Amit P. Sheth
• LSDIS Lab, Dept of Computer Science, University
  of Georgia

2
Outline

• Motivation
• Process Adaptation
• Empirical Evaluation
• Conclusions, Related Work and Future Agenda

3
Motivation

• Evolution of business needs drives IT innovation
• Initial focus on automation led to workflow
  technology
• In order to facilitate efficient
  inter-organizational processes distributed
  computing paradigms were developed
• CORBA, JMS, Web Services
• The current and future needs include
• Creating highly adaptive process that react to
  changing conditions
• Focus on real time events and data RFID and
  ubiquitous devices
• Have the ability to quickly collaborate with new
  partners
• Aligning business goals and IT processes

4
Motivation
Each enterprise will measure and aspire to its
own unique level of dynamism based on its
individual purpose. It is about being nimble and
adaptable. A fully integrated business platform
can respond faster, and completely, to change.
Whether it involves fulfilling a new mandate or
embracing a new market opportunity. Some
organizations will push the envelope, automating
event-triggered responses for highly integrated
closed-loop processes, setting the stage for
self-optimizing systems. Sandra Rogers, White
Paper Business Forces Driving Adoption of
Service Oriented Architecture, Sponsored by SAP
AG

• Current Tools focus on allowing businesses to
  have greater dynamism and agility
• Microsoft Dynamics, IBM Websphere Business
  Integration, SAP Netweaver
• All of these Current focus on dynamic and agility
  through human interaction using GUIs
• All of them list SOA (WS) as a technology for
  realization
• The future
• Move focus to greater automation
• Capture domain knowledge and declaratively
  specify criteria for process configuration
  (Dynamic process configuration)
• Add decision making support to process execution
  tools for process adaptation (Process Adaptation)

5
SOA Maturity Model
Adaptive/Autonomic
6
Levels of autonomic maturity
System monitors, correlates and takes action
Dynamic Business policy based management
Established Standards
Correlation and guidance
Centralized tools and manual analysis
Manual Analysis
No Established Standards
7
Motivating Scenario

• Consider a simplified supply chain process of a
  computer manufacturer
• Most parts are multiple sourced (overseas and
  internal suppliers)
• Suppliers characterized as preferred or secondary
• Overseas goods cheaper but greater lead times
• There often exist part compatibility constraints
• Choosing a certain motherboard restricts choices
  of RAMs, processors
• Usually important to maintain production schedule
  in the presence of delayed orders

8
Process Adaptation

• Ability to adapt the processes to external events
• Expected events
• Unexpected events
• Two kinds of failures
• Failures of physical components like services,
  network
• Can replace services using dynamic configuration
• Logical failures like violation of SLA
  constraints/Agreements such as Delay in delivery,
  partial fulfillment of order
• Need additional decision making capabilities

9
Process Adaptation
Adaptation Problem Optimally adapt to events like
delays in ordered goods

• Conceptual Approach
• Maintain states of the process normal states,
  error states, goal states
• Capture costs while transitioning from error
  states to goal state
• Ability to decide optimal actions on the basis of
  state

10
Process Adaptation

• Research Challenges
• Creating a model to recover from failures and
  handle external events
• Model must deal with two important factors
• Uncertainty about when a failure occurs
• Cost based recovery
• Scenario
• After order for MB and RAM are placed, they may
  get delayed
• The manufacturer may have severe costs if
  assembly is halted
• It must evaluate whether it is cheaper to
  cancel/return and reorder or take the penalty of
  delay
• Caveat possible that reordered goods may be
  delayed too

11
New Framework

• Introduce a framework within which to study
  process adaptation
• Two criteria
• Cost-based optimality
• Computational Efficiency

Decentralized Adaptation
Centralized Adaptation
Hybrid approaches
Decreasing Optimality
Decreasing Computational Efficiency
12
High Level Architecture
Entities Process Manager (PM) Responsible for
global process configuration Service Manager
(SM) Responsible for interaction of process with
service Configuration Module (CM) Discovery and
constraint analysis Adaptation Module (AM)
Process adaptation from exceptions/events
13
Modeling Decision Making Process of Service
Managers using MDPs

• Each Service Manager is controlled by a MDP
• SM ltS, A, PA, T, C, OCgt , where
• S is the set of local states of the service
  manager.
• A is the set of actions of the service manager.
  The actions include invoking Web service
  operations and calling the configuration
  manager.
• PA S ? A is a function that gives the
  permissible actions of the service manager from a
  particular state.
• T S A S ? 0, 1 is the local Markovian
  transition function. The transition function
  gives the probability of ending in a state j by
  performing action a in state i.
• C S A ? R is the function that gives the cost
  of performing an action from some state of the
  service manager.
• OC is the optimality criterion. We minimize the
  expected cost over a finite number of steps, N,
  also called the horizon.

14
Policy Computation

• The optimal action at each state is represented
  using a policy.
• In order to compute the policy, a value is
  associated to each state.
• The value represents long term expected cost of
  performing the optimal action from that state and
  is calculated the following dynamic programming
  equation.

The policy pi S N ? R is then computed as
N is the number of steps to go and Gamma is the
discount factor Algorithm developed by Bellman in
57
15
Marginalizing events
16
Generating States using preconditions and effects
Actions
Chance Variables
Events
17
Generated State Transition Diagram
State No. Values of Boolean variables Explanation
1 Ordered
2 Ordered and Canceled
3 Ordered and Delayed
4 Ordered, Received and Returned
5 Ordered, Delayed and Cancelled
6 Ordered, Delayed, Received and Returned
7 Ordered, Delayed and Received
8 Ordered and Received
18
Costs and Probabilities

• Costs of ordering taken from configuration module
• From first two service sets
• Optimal supplier and alternate supplier
• Probability of delay and cost of returning and
  canceling taken from supplier policy
• Can be represented using WS-Policy or
  WS-Agreement

19
Supplier Policy

• The supplier gives a probability of 55 for
  delivering the goods on time.
• The manufacturer can cancel or return goods at
  any time based on the terms given below.
• If the order is delayed because of the supplier,
  the order can be cancelled with a 5 penalty to
  the manufacturer.
• If the order has not been delayed, but it has not
  been delivered yet, it can be cancelled with a
  penalty of 15 to the manufacturer.
• If the order has been received after a delay, it
  can be returned with a penalty of 10 to the
  manufacturer.
• If the order has been received without a delay,
  it can be returned with a penalty of 20 to the
  manufacturer.

20
Costs and Probabilities
21
Handling Coordination Constraints

• Since the RAM and Motherboard must be compatible,
  the actions of service managers (SMs) must be
  coordinated
• For example, if MB delivery is delayed, and MB SM
  wants to cancel order and change supplier, the
  RAM SM must do the same
• Hence, coordination must be introduced in SM-MDPs

22
Centralized Approach

• State space created by Cartesian product of
  transition diagrams
• Joint actions from each state
• Transition probability created by multiplying
  states
• Costs created by adding cost per action from each
  state
• Compatible actions given rewards
• Incompatible actions given penalties
• Optimal but exponential with number of manager

23
Decentralized Approach

• Simple coordination mechanism
• If one service manager changes suppliers
• All dependent managers must change suppliers
• Low complexity but sub-optimal

24
Hybrid Approach

• If the policy of some SM dictates it to change
  suppliers, the following actions happen
• it sends a coordinate request to PM
• PM gets the current cost of changing suppliers
  or current optimal action by polling all SMs
• It takes the cheapest action (change supplier or
  continue)
• A bit like decentralized voting- will change
  suppliers if majority are delayed
• It mirrors performance of centralized approach
  and has complexity like the decentralized approach

25
Evaluating Process Adaptation

• Evaluation with the help of the supply chain
  scenario
• Two main parameters used for the evaluation
• Probability of Delay (probability that an item
  ordered from a supplier will be delayed)
• Penalty of Delay (cost for the manufacturer for
  not reacting to delay)
• Total process cost 1000 and cost of changing
  suppliers (CS) 200

26
Evaluating Adaptation


KEY M-MDP Centralized Random Random process
(changes suppliers for 50 of delays) Hyb. Com
Hybrid MDP-Com Decentralized
27
Evaluating Adaptation


28
Evaluating Adaptation


29
(No Transcript)
30
Observations

• Results
• For Penalty 200 (cost of CS cost of delay),
  MDP always waits
• For Penalty 300, 400 (cost of CS lt cost of
  delay), MDP changes at lower prob., waits at
  higher prob.
• Conclusions
• Thus MDP makes intelligent decisions and
  outperforms random adaptation that changes
  suppliers 50 of the time it is delayed
• Centralized MDP performs the best, followed by
  Hybrid MDP

31
Related work

• Focus on correctness of changes to control flow
  structure
• Adept1, Workflow inheritance 2, METEOR
• Use of ECA rules 3 to automatically make
  changes
• Change of service providers based on migration
  rules in E-Flow 4
• We extend previous work in this area by using
• Cost based adaptation
• Coordination Constraints across services

1 M. Reichert and P. Dadam. Adeptflex-supporting
dynamic changes of workflows without losing
control. Journal of Intelligent Information
Systems, 10(2)93129, 1998 2 W. van der Aalst
and T. Basten. Inheritance of workflows an
approach to tackling problems related to change.
Theoretical Computer Science, 270(1-2)125203,
2002. 3 R. Muller, U. Greiner, and E. Rahm.
Agentwork a workflow system supporting
rule-based workflow adaptation. Journal of Data
and Knowledge Engineering, 51(2)223256,
2004. 4 Fabio Casati, Ski Ilnicki, Li-jie Jin,
Vasudev Krishnamoorthy, Ming-Chien Shan Adaptive
and Dynamic Service Composition in eFlow. CAiSE
2000 13-31
32
Conclusions and Future Work

• Showed the utility of Markov Decision Processes
  for optimal adaptation of Web processes
• Adaptation is need to handle logical failures and
  events
• Whether to adapt or not depends on the cost of
  the failure
• For this evaluation it was the cost of the delay
• In the real world things often go wrong or not as
  expected
• Earlier processes were static or real time events
  were not available as easily
• Many researchers/industry vendors seeking to
  create adaptive business process frameworks
• This is one of the first works that provides cost
  based adaptation
• Future Work
• Move towards autonomic Web processes