Challenges in Adapting Automated Planning for Autonomic Computing

1
Challenges in Adapting Automated Planning for
Autonomic Computing

• Biplav Srivastava Subbarao Kambhampati
• IBM India Research Lab Arizona State
  University
• sbiplav_at_in.ibm.com rao_at_asu.edu
• ICAPS 2005, Monterey, CA, USA
• (Also being presented at 2nd Intl. Conference on
• Autonomic Computing)

2
The Case for Automated Planning in Autonomic
Computing

• Biplav Srivastava Subbarao Kambhampati
• IBM India Research Lab Arizona State
  University
• sbiplav_at_in.ibm.com rao_at_asu.edu
• ICAC 2005, Seattle, USA
• Presented by Hemal Khatri

3
Planning in Autonomic Computing (AC)

• The P of the M-A-P-E loop in an Autonomic
  Manager
• Planning provides the policy engine for goal-type
  policies
• Given expected system behavior (goals), determine
  actions to satisfy them
• Synthesis, Analysis Maintenance of plans of
  action is a vital aspect of Autonomic Computing
• Example 1 Taking high-level behavioral
  specifications from humans, and control the
  system behavior in such a way as to satisfy the
  specifications
• Change requests (e.g., INSTALL, UPDATE, REMOVE)
  from administrator in managing software on a
  machine (Solution Install scenarios)
• Example 2 Managing/propagating changes caused by
  installations and component changes in a
  networked environment
• Remediation in the presence of failure

Autonomic Manager
Plan
Analyze
Knowledge
Monitor
Execute
Managed Element
4
Information Expected to be Available while
Planning in AC Scenarios

• Planning is ltP, I, G, Agt
• P is a set of predicates
• I and G are initial and goal states drawn from P
• A is a set of actions, Ai with
• Aipre (preconditions) Aipost (postconditions)
  drawn from P

Scenario I (initial state) G (goal state) A (actions) S (existing plans) Constraints (domain constraints)
Self-configuring Yes Yes - - Yes
Self-healing Yes Yes Yes - Yes
Self-optimizing - - - Yes Yes
Self-protecting - Yes - Yes Yes
5
Comparing Current Status of Automated Planning
and the Needs of AC planning

• Highly scalable planners exist for synthesizing
  plans of actions. However
• They expect complete domain theories
• They focus on plan generation rather than plan
  management

• Planning technology is relevant for AC computing,
  but we also need
• Ability to handle incomplete domain theories
• Focus on plan management rather than just plan
  synthesis
• Support mixed initiative continual (re)planning

Early systems in AC a) CHAMPS Domain-dependent
planner for self-configuration b)
ABLE-Planner4J Domain-independent
planning for self- but expects complete
I,G, A.
6
Planning with Incomplete Domain Theories

• Domain theory is partial if correctness cannot be
  causally explained
• Domain theory ? Explanation
  ? Modification
• HTNs provide natural support
• Explainability Event vs. State constraints
• EVENT If you do a, then do b before c (dont ask
  why!)
• STATE The condition p is required by a and is
  given by b
• State constraints can be compiled to event
  constraints. But the reverse?

• In Autonomic computing (as well as web-service
  composition, scientific workflow handling), the
  planner doesnt have access to complete and
  correct specification
• Action specifications may be incomplete
• Domain theory may be in terms of dependencies
• The planner cant always verify correctness
• ..but can certainly look for errors in a plan

7
Prescriptions

• AC Practioners
• Leverage current planning solutions in convenient
  scenarios very efficient and will answer qns
  such as
• What interactions will occur if a new operation
  is introduced into the plan
• What high-level goals will go unsupported if an
  action is removed
• Expend time in effect-based modeling
• Complete specifications make it easy to provide
  the causal dependency structure of the plan. This
  in turn helps in plan-management by allowing us
  to answer questions such as
• What interactions will occur if a new operation
  is introduced into the plan
• What high-level goals will go unsupported if an
  action is removed

• Planning Researchers
• In AC, we can at most expect incomplete
  specification
• Ordering constraints may be provided without an
  explanation of why they are needed
• Some information about incompatibility of actions
  may be provided
• Managing such plans poses two technical
  challenges
• Deriving additional dependencies between workflow
  operations
• Adapting planning techniques to deal with partial
  causal information

8
Summary

• We developed an understanding of the planning
  needs of AC computing
• Connections with 2 other very close
  applicationsWeb Services, and Scientific
  Workflow management
• Evaluated the match between existing planning
  technology and AC computing needs, and identified
  specific needed extensions
• Currently focusing on plan synthesis and
  management with incomplete domain theories (such
  as are present in AC computing scenarios)
• Impact will be measured in terms of availability
  of information sought about the domain and
  improvement in the quality of plans handled
  (analyzed/ generated/ managed).
• Benchmarking will be in the software installation
  and problem determination scenarios.