17-811 Self-Healing Systems: Class Discussion Summary

Summary of Class Discussion for April 23 by Kevin Steppe
 

Agent based approaches

* necessary/ required attribute
(others are typical but 'optional')
*autonomous and task based
                - asynchronous from the user interactions
*communicates w/ other agents & env
                 - tend to be very focused on small tasks
                 - need to work w/ other agents to finish larger tasks
        reactive and learning
                -
        goal-oriented and pro-active
 
        mobility/flexibity
                - move around nodes of a network
                        able to move to a conducive environment for the tas
 
Benefits --
        decomposed task in bottom up building
        cooperative redundancy
                - agents can have overlapping functionality for specialization & redundancy
        Viewpoints
                - working at multiple levels of abstraction
        Quasi-blackboard architecture
                - ind. running components solving small problems can together solve larger problems

Domains --
        Distributed, UI, Collaboration, Games
 
Tools --
        Agents, pre/post processing
                - pre-processing can partly be implemented by bidding on taking action
 
Goals --
        Dependability
        Performance
        Service Modularity
        Emergent Behavior generating complex Services

        - How do we guarentee that the system can adapt and evolve?
                What prevents it from getting stuck?
 
 
--- Robust Software Paper ---
        Independant agents can be guided or controlled by a higher level
        They then communicate to solve larger problems
        Compenents can be hidden if tasks are well defined
 
        Similar to N-version programming w/ similar problems
        Designing an ontology for service interaction is a huge problem
                - they claim this isn't needed for agents
                - then how do they communicate?
                -> break the large ontology into many smaller communication agreements
                        but it's not clear that the smaller agreements are easier
 
        Soccer Robots are an example of an agent system
 
---------------------------
Analogies to Biological systems
        - cells communicating locally
        - local adaptation
 
        - encapsulate well understood human decisions
 
 
---- Investigation into Self-Adaptive Software Agents Development ---

Framework to use adaptive agents

System 1 - Operations
                - Operate the motor to move limbs
                - optimize and report on power, etc.
System 2 - Coordination
                - coordinate agents at level 1
                - coordinate all the limbs to move smoothly
                - prevent oscillation behavior
System 3 - Control  & Auditing operations
                - are the actions in line with the goals
                - resource usage planning
System 4 - Environment and oportunity

System 5 - Desires & world model
 

Diversion -- Learning cognitive assistant

        Agent to help improve performance in 4 tasks
        * Email assistant
                1. categorization on input stream
                        - it's a meeting
                2. info extraction via text understanding
                        - what does the email say about this meeting?
                3. Agent
                        - meeting planner
                        - course assistant
                        - etc.
        * Scheduling
        * Web site management
                - project page updates, etc.
        * Crisis management
                - multi-user coordination
 

Aura view:
                User          ->        Task
        -----------------   /   ---------------------
    --->|       App     | -/    |       Env Manager |           Task Descriptions
        -----------------       ----------------------
        |       OS/FS   |       |                   |
        -----------------       ----------------------
 
Filtering/Categorization is moving email from app level to task level
Here, the environment (service availability) is not variable, we don't need Env. Manager
If there are multiple App choices to carry out the task, may require a dynamic mapping

Learning based on how the user acts with the apps, to discover preferences / categorization / etc

Context-aware computing

Wearable Computing

Exploratory Systems
Customer Systems
Visionary Design/Research

Rate of change of data
        - Text (weekly)
        - Work orders (daily)
        - collaboration (seconds/minutes)

Platform for incrementing base on field results

Expanding abilities
        Representation
        Input
        Output

Collaboration improvement: prestored expertise, Master/App, Teaml, Synthetic

Evolutionary design
        Prestored
        -> add in Synchronous Shared Space for Help desk
        -> add Asynchronous Q/A space from team
                - Audio BB, etc.
        -> Proactive Assistant
                - Synthetic helper (FAQ - interview)
                - Context Aware (cognitive model / overload, etc)

Declarative Memory
        --> Production Rules
        --> Production Rules' Actions
 
Build Context from 20 attributes
        Location -> Social geography -> ....
        Physical response -> emotional state -> cognitive load ...
 
        Context awareness builds from physical setting