agents: pros and cons keita fujii jennifer rhough
Post on 21-Dec-2015
216 views
TRANSCRIPT
PapersAgents that Reduce Work and Information Overload (P. Maes, p. 525-536) Presenting Through Performing: On the Use of Multiple Lifelike Characters in Knowledge-Based Presentation Systems (E. André, T. Rist; IUI-2000, pp. 1-8) Embedding Critics in Design Environments (G. Fischer, 537-561) Multimodal Interaction for Distributed Interactive Simulation (P. Cohen et al., 562-571) Animated Conversation: Rule-based Generation of Facial Expression, Gesture and Spoken Intonation for Multiple Conversational Agents (J. Cassell, J.et al., p. 582-591) Direct manipulation vs. interface agents§§ (B. Shneiderman, P. Maes; Interactions 4, 1997, p. 42-61)
These papers focus on agents that:Support task performance Perform tasks on behalf of users Present informationEnable integration of complex software
systemsCreate interfaces possessing
anthropomorphic communicative abilities integrated speech, facial displays, gesture
Agents that Reduce Work and Information Overload
MotivationWe increasingly use computers for our
everyday activities Increasing number of untrained users
Dominant Interaction MetaphorDirect manipulation vs. Indirect
management
Building Agents
Two problems to overcomeCompetence
How, when and whatTrust
Comfort levels in delegating tasks Integrate into existing interfacesWay of operating should be easily
understandable
Past Approaches
“Semi-autonomous Agents” Example: Email sorting
agent Consists of a collection of
user programmed rules Competence not dealt
with Trust
Do you trust your own skills?
“Knowledge-based approach” Interface agent supplied
with extensive domain specific knowledge
Competence issues Trust issues
Another approach
Hypothesis is that under certain conditions the agent can “program itself”
Two conditions need to be fulfilledUse of application involves repetitive
behaviorThis behavior is potentially different for all
users
Personal Assistant Metaphor
Assists user by: Hiding complexity of difficult tasks Performs tasks on user’s behalf Trains/teaches user Helps different users collaborate Monitors events and procedures
The learning approach
Requires less work from the end-user and application developer
Is a solution to the trust problem Allows agents to reason their behavior Agents can more easily adapt to the user over
time and become customized to individual/organizational preferences and habits
Helps in transferring info, habits, and know-how among the different users of a community.
Electronic Mail Agent (Maxims)
Learning technique is memory-based learning Learns to prioritize, delete, forward, sort, and
archive mail on behalf of user by “looking over the shoulder” of the user
Agent memorizes generated situation-action pairs Situations described by features
Maxims
Agent will compare new with memorized situations and tries to find a set of nearest neighbors to base its action distance metric – weighted sum of differences for
the features; weight determined by agent agent analyzes its memory for correlation bet
features and actions taken From vs Date
measures confidence in prediction
Maxims
Slow start problem user can instruct agent explicitly
default or hard-and-fast rules, use of “wildcard” fields Multi-agent collaboration
Confidence is below “tell-me” so ask other agents by sending part of description via email
Learns trustworthy sources
Preliminary user approval Report feeling comfortable delegating tasks Users want to be able to instruct agent to
disregard behavior
Meeting Scheduling Agent
Same software agent as above but attached to a meeting scheduling package assists user in
scheduling of meetings (accept/reject, schedule, reschedule, negotiate times)
News Filtering Agent
User creates “news agents” and initialize by giving it +/- examples of articles User can give feedback on portions of articles recommended No social filteringLimitations Users rely on it too much - still responsible for
finding less predictably interesting articles Restriction to keywords only, no semantic analysis
Entertainment Selection Agent
Social filtering Relies solely on correlations between different users
Problems Users can rely too much and not enter new
information on items discovered themselves How to jumpstart the system so agents notice
correlations Users can rely too much and not enter new
information on items discovered themselves
Virtual Users
Some questions to ask
How to guarantee user’s privacy?
How can heterogeneous agents collaborate?
Should the user be held responsible for the agent’s actions?
Presenting through Performing: On the Use of Multiple Animated Characters in Knowledge-Based Presentation
Systems
Based on observation that vivid and believable dialogues are a means to present information to an audience
Use of animated characters Ability to express emotions in a believable way Provide means of conveying conversational
signals Users rate presentations by characters as lively
and engaging
Rationale
Presentation teams vs. face to faceEasier to convey differing points of view
Debates between two charactersAllows reinforcementSingle members function as indices to help
user classify informationAlso used to convey meta-information
Some people feel uncomfortable when addressed directly by an agent
Related Work
Virtual human-like weather reporter One agent for presenting information
Bank teller and employee Restricted to Q&A type dialogue between two
agents
Mr. Bengo Resolutions of disputes with judge, prosecutor,
and lawyer (controlled by user) Exhibits basic emotions but not through linguistic
style
Designing Presentation Dialogues
Choose dialogue typeSales dialogue and soccer commentary
Define rolesSales – seller and buyer
Define characters to occupy rolesPersonality and emotional traits
Gestures, linguistic styleDistinguishable by expertise, audio/visual
appearance, interests
Generation of Dialogue
Actors with scripted behaviors Actors in a play Knowledge to be communicated known a priori Possible to vary dialogue by expressions,
gestures, emotions
Autonomous actors Agents draw from dialogue strategies to meet a
certain goal (can be different) Reactive and difficult to ensure coherence
Inhabited Market PlaceScriptedOrdinary product database each example with
n attributes
Attributes also grouped according to the values of the character safety, economy,
comfort, prestige, environmental considerations, etc
Design of Information Dialogues
Central planning component
Knowledge is represented by plan operators handle the dialogue
and allocation of dialogue agents.
NAME: “DiscussValue1”
GOAL: PERFORM DiscussValue $attribute;
PRECONDITION:
FACT polarity $attribute $dimension “neg”;
FACT difficulty $attribute $dimension “low”;
FACT Buyer $buyer;
FACT Negative $buyer;
FACT Seller $seller;
BODY:
PERFORM NegativeResp $buyer $dimension;
PERFORM RespNegResp $seller $attribute $dimension;
Generation Example
Agent Role Personality factors Interests
Robby seller extravert, agreeable sportiness
Peedy buyer introvert, disagreeable environment
Peedy: How much gas does it consume?
Robby: It consumes 8 liters per 100 km.
Peedy: Isn’t that bad for the environment?;;;negative comment because it is disagreeable, less direct speech
;;;because it is introvert
Robby: Bad for the environment? It has a catalytic converter. It is made of recyclable material.
;;;questions the negative impacts and provides counter arguments
RoboCup Soccer Games
Semi-autonomous agents triggered by events occurring in the scene
& dialogue from other agent rapidly changing environment
Gerd and Matze Characterized by sympathy of team, level
of extraversion, openness, and two emotional dispositions, excitability, and valence
Dialogue Input and Templates
Basic input is obtained by the soccer server delivers player location and orientation, ball
location, score, play modes (goal kicks, throw-ins) info is pieced together at a more conceptual level to
provide material for characters
Templates extracted from 13.5 hours of actual soccer reports and characterized by features like verbosity, bias, formality Selection of template filtered by
situational needs like time remove templates that were recently used keep those that are aligned with character’s attitude keep those aligned with character’s personality
Generation ExampleAgent Attitude Personality factors
Gerd in favor of team Kasunga extravert, open
Matze neutral introvert, not open
Gerd: Kasunga kicks off;;;recognized event: kick off
Matze: Andhill 5;;;recognized event: ball possession, time pressure
Gerd: We’re live from an exciting game, team Andhill in red versus Kasunga in yellow
;;;time for background information
(…)
Gerd: ball hacked away by Kasunga 4;;;recognized event: shot, flowery language since it is creative
ConclusionsTesting Users found the scenarios entertaining and
amusing Eager to test the effect of role castings on the
generated presentation Implies people might learn more about a subject matter
because they are willing to spend more time with a system
Questions How to actively involve the user, either as a co-
presenter or by providing feedback during performance
Optimal number of roles and casting
Embedding Critics in Design Environments
The critiquing approach Growth of human knowledge Helps in error elimination Promotion of mutual understanding of all
participants
Computer based critiquing applied to design Critics recognize and communicate debatable
issues Suited for design tasks where
Knowledge of design domain is incomplete/evolving Design knowledge is distributed Problem requirements can only be partially specified
Shortcomings that hinder the ability to say the “right” thing at the “right” time
Lack of domain orientationInsufficient facilities for justifying critic suggestionsLack of an explicit representation of user’s goalsNo support for different perspectivesTiming problemsPassive vs. active critics
HYDRA-KITCHEN
Design creation tools Construction component
Analogous to the Paint program Includes palette of domain-oriented design units (e.g.
sinks, stoves) Critics are tied to units and relationships between units
Specification component Allows designers to describe abstract characteristics of
their design Dynamic
Used to tailor critic’s suggestions and explanations
HYDRA-KITCHEN IIDesign information repositoriesArgumentative hypermedia component
Consists of issues, answers, and arguments about decisions made in the design
Identifies pros and cons of a suggestion and helps users to understand consequences of following a suggestion
Catalog componentCollection of previously constructed designsCan be used by critics as examples illustrating
solutions
Generic Critics
Enabled by placing design units into the construction area
Reflects knowledge that applies to all designs
Defined through property sheets that specify rules and relations Users can add and modify
Specific critics
Enabled by the partial specification
Fine tune generic critics
Detects inconsistencies between design specification and construction Situation specific physical characteristics
Size/shape of kitchen, owner’s height
Specified requirements Abstract domain concepts like safety or efficiency
Interpretive critics
Enabled by the currently active design perspective
Examines the design from different viewpoints Electrician, plumber, city inspector, interior
designer
Inheritance network - inherit other critics
Can then add additional rules and modify inherited ones
Some advantages
Embedding allows access the work state and time delivery of information that is relevant to the current taskSupport for different perspectivesCritic suggestions supported by domain-oriented design environmentDesign environment allows explicit representation of user’s goalsLocating relevant information Large information space
People are lazy or unaware