agents that reduce work and information overload and beyond intelligent interfaces presented by...
TRANSCRIPT
Agents that Reduce Work and Information Overloadand
Beyond Intelligent Interfaces
Presented byPresented by
Maulik OzaMaulik OzaDepartment of Information and Computer ScienceDepartment of Information and Computer Science
University of California, IrvineUniversity of California, Irvine
[email protected]@ics.uci.edu
ICS 205 – Spring 2002ICS 205 – Spring 2002
Agents that Reduce Work and Information Overload
Pattie MaesPattie Maes
Why Agents?
Computers assisting in everyday tasksComputers assisting in everyday tasks Untrained users interacting with computersUntrained users interacting with computers Computers require continuous user interactionComputers require continuous user interaction ““Indirect Management”Indirect Management” required instead of required instead of
“direct manipulation”“direct manipulation” CollaborationCollaboration with the user as a “personal with the user as a “personal
assistant”assistant”
Figure: The interface agent does not act as an interface or layer between the user and the application. Rather, it behaves as a personal assistant which cooperates with the user on the task. The user is able to bypass the agent.
Agents Duties’
Perform tasks on the users behalfPerform tasks on the users behalf e.g. Selection of bookse.g. Selection of books
Train or teach the userTrain or teach the user e.g. Image editinge.g. Image editing
Help different users collaborateHelp different users collaborate e.g. Meeting schedulinge.g. Meeting scheduling
Monitor events and procedures on the user’s Monitor events and procedures on the user’s behalfbehalf e.g. Information filteringe.g. Information filtering
Building Agents – Problems
CompetenceCompetence How does an agent acquire the knowledge?How does an agent acquire the knowledge?
TrustTrust How does the user feel confident delegating How does the user feel confident delegating
the task?the task?
Previous Approaches End user program the interface agentEnd user program the interface agent
User programmed rulesUser programmed rules DisadvantagesDisadvantages
Does not deal with the competence criterionDoes not deal with the competence criterion Requires too much insight from the end userRequires too much insight from the end user
Knowledge-based approachKnowledge-based approach Domain knowledge programmed into the agentDomain knowledge programmed into the agent DisadvantagesDisadvantages
Work for programming the knowledgeWork for programming the knowledge Adaptation to particular users preferencesAdaptation to particular users preferences Trust a big issueTrust a big issue
Approach – Machine Learning
Under Under certaincertain conditions the agent program itself conditions the agent program itself Limited background knowledgeLimited background knowledge Learns from user and other agentsLearns from user and other agents
Conditions for the agent to learnConditions for the agent to learn Repetition an important aspectRepetition an important aspect Behavior different for all usersBehavior different for all users
The metaphor – “personal assistant”The metaphor – “personal assistant” Learns based on the preferences of the employerLearns based on the preferences of the employer Requires time for performing efficientlyRequires time for performing efficiently Learns based on experience, employer’s instructions as Learns based on experience, employer’s instructions as
well as from experienced assistantswell as from experienced assistants
Advantages of the Approach
Less workLess work AdaptationAdaptation Transferring InformationTransferring Information
Learning Technique
Observe and imitateObserve and imitate Adapt based on user feedbackAdapt based on user feedback
Direct feedbackDirect feedback Indirect feedbackIndirect feedback
Trained based on examplesTrained based on examples Advice from other agentsAdvice from other agents
Figure: The interface agent learns in four different ways: (1) it observes and imitates the user's behavior, (2) it adapts based on user feedback, (3) it can be trained by the user on the basis of examples, and (4) it can ask for advice from other agents assisting other users.
Example agents
Electronic mail handling agentElectronic mail handling agent Meeting scheduling agentMeeting scheduling agent Electronic news filtering agentElectronic news filtering agent Recommending agentRecommending agent
Electronic Mail Agent – Maxim
Learns to prioritize, delete, forward, sort and archive Learns to prioritize, delete, forward, sort and archive mailmail
Uses Memory-based reasoningUses Memory-based reasoning Measures confidence level in the predictionMeasures confidence level in the prediction Actions determined by thresholdsActions determined by thresholds Dealing with initial low competenceDealing with initial low competence
Figure: Simple caricatures convey the state of the agent to the user. The agent can be "alert" (tracking the user's actions), "thinking" (computing a suggestion), "offering a suggestion" (confidence insuggestion is above "tell-me" threshold), "surprised" if the suggestion is not accepted, "gratified" if the suggestion is accepted, "unsure" about what to do in the current situation (confidence below "tell-me" threshold, and thus suggestion is not offered), "confused" about what the user ends up doing, "pleased" that the suggestion it was not sure about turned out to be the right one after all, and "working" or performing an automated task (confidence in prediction above "do-it" threshold).
Other Agents Meeting Scheduling AgentMeeting Scheduling Agent
Generic learning agent adapted to the scheduling software.Generic learning agent adapted to the scheduling software. News Filtering Agent – NewTNews Filtering Agent – NewT
Filter Usenet newsFilter Usenet news Agents can be trained for specific purposesAgents can be trained for specific purposes
Entertainment Selection Agent – RingoEntertainment Selection Agent – Ringo The “killer app”?The “killer app”? How to make enough data available to the system for it to How to make enough data available to the system for it to
make recommendationsmake recommendations User may rely too much on the system and stop entering new User may rely too much on the system and stop entering new
itemsitems Solution – “virtual users”Solution – “virtual users”
Beyond Intelligent Interfaces: Exploring, Analyzing, and Creating Success Models
of Cooperative Problem Solving
Gerhard FischerGerhard Fischer
Brent ReevesBrent Reeves
Cooperative Problem Solving
Augmenting a person’s ability to create, Augmenting a person’s ability to create, reflect, design, decide and reasonreflect, design, decide and reason
Conceptual framework behind a system Conceptual framework behind a system determines its behaviordetermines its behavior
Empirical Study
Study of a success modelStudy of a success model Highlights the inherent difficulties in high Highlights the inherent difficulties in high
functionality systemsfunctionality systems Necessary to get a better understanding of Necessary to get a better understanding of
the systemthe system
Results from the study (1/2)
Users do not know the existence of toolsUsers do not know the existence of tools Users do not know how to access toolsUsers do not know how to access tools Users do not know when to use the toolsUsers do not know when to use the tools Users cannot combine or adapt tools for Users cannot combine or adapt tools for
special usesspecial uses
Results from the study (2/2)
Incremental problem specificationIncremental problem specification Identifying the problemIdentifying the problem
Achieving shared understandingAchieving shared understanding Identifying the solutionIdentifying the solution
Integration between problem setting and Integration between problem setting and problem solvingproblem solving Context important in determining the Context important in determining the
problemproblem
Analysis based on the results
Natural Language is less important than Natural Natural Language is less important than Natural CommunicationCommunication
Multiple specification techniqueMultiple specification technique Mixed initiative dialoguesMixed initiative dialogues Management of troubleManagement of trouble Simultaneous exploration of problem and solution Simultaneous exploration of problem and solution
spacesspaces Humans operate in the physical worldHumans operate in the physical world Humans make use of distributed intelligenceHumans make use of distributed intelligence
Requirements for a Cooperative Problem Solving System
Beyond user interfacesBeyond user interfaces Problems in the contextProblems in the context Reliability of “Back talk” in design situations must be increasedReliability of “Back talk” in design situations must be increased Need for specialization and putting knowledge in the worldNeed for specialization and putting knowledge in the world Supporting human problem-domain communication with Supporting human problem-domain communication with
domain-oriented architecturesdomain-oriented architectures
Conclusions
Interfaces of the futureInterfaces of the future IntelligentIntelligent Context awareContext aware TrustworthyTrustworthy CompetentCompetent InvisibleInvisible
IssuesIssues PrivacyPrivacy EthicalEthical