Intelligent agents on the Web Adina Magda Florea
1. Information agents
Several types of information agents Personal agents • provide "intelligent" and user-friendly interfaces • observe the user and learn user’s profile • sort, classify and administrate e-mails, • organize and schedule user's tasks • in general, agents that automate the routine tasks of the users Web agents • Tour guides Search engines • Indexing agents - human indexing • FAQ finders - spider indexing • Expertise finders
Cooperative information retrieval systems w Use information retrieval theory and AI w Make information resources available by wrapping them with agents capabilities w Every agent is expert with its own repository w Agents communicate using an ACL
RETSINA: Reusable Environment for Task-Structured Intelligent Networked Agents w RETSINA is a domain-independent and reusable infrastructure on which MAS systems, services, and components live, communicate, and interact. w RETSINA is an architecture for developing distributed intelligent software agents that cooperate asynchronously to perform information management: information gathering, information filtering, information integration w RETSINA is project developed at the Robotics Institute, CMU
RETSINA MAS architecture
The agent architecture
Operating environment w Platform independent: – any platform that runs Windows, Linux, or Sun OS – Palm. Pilots w Multiple language support: Java, C/C++, Python, LISP, and Perl w Network support: TCP/IP, wireless, infrared, and serial connections
Web. Mate – an information search agent in RETSINA w Web. Mate is a personal agent for WWW browsing that enhances searches and learns user interests. w Information searching: § trigger pair model § document similarity based on relevance feedback
w Trigger Pair Model § If a word S is significantly correlated with another word T, then (S, T) is considered a trigger pair with S being the trigger and T being the triggered word w Relevance feedback § The user identifies relevant pages from an initial list of retrieved documents § the system analyzes the page using the context of keyword (i. e. the words near by) § the system finds out the relevant keywords § enlarge the user query using the relevant keywords
Information agents for e-communities (BTexact Technologies) Personal Agent Framework (PAF) w central profile management agent w suite of application agents that use profiles in conjunction with several information sources w Web-based agents
Profiler Agent w one for each user w stores interest information in a hierarchy in which interests lower in the hierarchy inherit their parent interest characteristics w transparent for the user w each interest: § private § restricted § public
Application agents w Bugle – uses profile information to generate a daily newsletter that contains articles relevant to the user’s interests w Grapewine – works in the background, periodically notifying members via email about other members who have similar interest profiles. i. Vine lets the user interactively locate members with similar interests. Shows the shared areas of interest so the use can decide. w Pandora – helps broaden user’s interests via collective filtering, suggests new interests for members to explore. w Radar – just-in-time information agent; monitors the user current activity while, for example, authoring a document, and offers relevant information resources, news reports, FAQs; allows interaction with i. Vine.
Architecture w PAF is integrated with Open Agent Middleware, a Java virtual machine platform developed at Fujitsu Laboratories Evaluation w BTexact – internal trials w External trial at Concert – global venture between BT and AT&T w 1 community – around 1000 members
2. Agents for e-learning
Agent’s roles in e-learning w Enhance e-learning content and experience § give help, advice, feedback § act as a peer learning § participate in assessments § participate in simulation § personalize the learning experience w Enhance LMSs § facilitate participation § facilitate interaction § facilitate instructor’s activities
ADELE w Pedagogical agents developed by Center for Advanced Research in Technology for Education (CARTE) at USC / ISI to assist students in working through course materials w The lead character, an agent named Adele (Agent for Distance Learning Environments), is a pedagogical agent designed to work with Webbased educational simulations.
w Adele consists of a pedagogical agent and a 2 D animated persona, which is implemented as a web-based Java applet. w Adele: § adapts the presentation of the material as needed § monitors student’s progress § provides feedback, hints and rationales to guide student actions § references relevant material § evaluates student performance by probing questions. w She is used in two medical education systems: case-based diagnosis and trauma care.
w Simulations created for the course in diagnostic skill development presents the student with actual cases, including patient history, results of exams, lab tests, x-rays, CT scans and other diagnostic imaging methods. w By questioning and examining the virtual "patient" and studying clinical data, the student is able to practice diagnostic skills.
w Trauma care is a collaborative activity - physicians and paramedics work with other emergency response personnel. w Adele functioning includes the notion of "situations". w A situation is a high-level description of an "interesting state" along with a description of steps to take in that situation. w The animated persona is a Java applet. It can be used alone or with a Web page-based Java. Script interface, or incorporated in larger simulations.
ADELE’s architecture Web browser Simulation engine and GUI Animated persona Text-to-speech engine Reasoning engine Task Planner, Assessor Adele client Web server Adele server Store for case, student, persona, references, and simulation parameters Architecture of single user system. In the multi-user system, RE is server-based, as is the Session Manager Student model, case task plan, initial state Student record of actions
Task representation w Task plan = task steps and their dependencies, step rationale w task steps = object-oriented data structures processed by Adele’s Java-based reasoning engine w Reasoning engine – runs in 3 modes § restricts unsolicited input; Hint, Why § practice mode; Hint § exam; Adele is not available w Situation – triggers a plan w Situation plans are pre-authored w Adele’s reasoning situation-monitoring task; situation-based reasoning.
Pedagogy w Situation-based reasoning allows the recognition of pedagogical opportunities § ask questions related to a particular task § give feedback to chosen answers § ask follow-up questions § give references significant to a particular task § verify correctness of plan step order § records the student’s actions § analyze student’s record and provides domain appropriate feedback (e. g. , evaluation of diagnosis, evaluation of diagnostic’s costs, evaluation of the steps taken).
Adele’s persona w Uses gaze and gestures to react to student’s actions – repertoire of facial expressions and body postures that represent emotions: surprise, disappointment, etc. w Senses user’s mouse pointing, turns her head and looks toward that point. w She has also a pointer that she can use to point to objects in other windows. w Animations are produced from 2 -dimensional drawings => makes possible to run on a variety of desktops (no 3 D graphics needed).
STEVE w Developed at Information Science Institute, USC w Learning environment: simulation of the naval training facility in Great Lakes, Illinois w Steve – a 3 D pedagogical agent w Training: a 3 D, interactive, simulation environment
w Students and Steve agents are immersed in the simulation environment w Students – 3 D immersive view of the virtual world through a head-mounted display (HMD) and interacts with the world via data gloves w Lockheed Martin’s Vista Viewer software uses data from a position and orientation sensor on the HMD to update the students’ view as he moves around w Additional sensors on the glove keep track of the students’ hands and Vista sends messages when the student touches virtual objects
Humans and agents communicate through spoken dialogue An agent speaks to a person by sending a message to the person’s text-to-speech software – broadcasts the utterance through the headphones mounted on the HMD w Entropic’s True. Talk for speech synthesis w Students speak to the microphone on the HMD - sends the utterance to the speech recognition software – semantic representation of the utterance to the agents. w Entropic’s Grap. Hvite for speech recognition
Steve’s cognitive architecture Task knowledge Abstract motor commands Pedagogical capabilities Perception snapshots important events Soar rules Motor Control Detailed motor commands Simulator Spatial properties Message Dispatcher Perception Relevant events Visual, audio effects Interface components
w Separation between domain independent capabilities and domain specific knowledge w Perception, cognition and motor control modules: general capabilities independent of a particular domain: § planning § replanning § plan execution § assessment of student’s actions § question answering (What should I do next? , Why? ) § episodic memory § communication § control of human figure
w Course author specifies the domain knowledge in a declarative language w Domain knowledge § perceptual knowledge: – knowledge about objects in the virtual world, objects’ simulation attributes and spatial properties § task knowledge: – procedures for accomplishing domain tasks and text fragments for talking w Tasks: set of steps § ordering constraints § causal links § hierarchical planning
w Steve – acts as a tutor or learning companion w Steve was extended to support team training w Steve agents can play two roles: § tutor for an individual team member § can substitute for missing team members
w Tasks were extended with roles for different participants w Planning is extended by mapping task steps to team roles: roles are assigned during plan creation w Team task request: § each Steve agent involved in the task as a team member or instructor uses his task knowledge to construct a complete task model § New types of actions - a speech act from one team member to another § each speech act appears as a primitive action in task description
Learning Companion that recognizes affect w MIT Media Lab w Affective states significant to learning: anxiety, worry/boredom, indifference, interest, curiosity, confident, etc. w on-goal and off-goal
Affect recognition w Posture w Eye-gaze w Facial expression w Hand movement
On Task Off Task Posture Leaning forward Slumping on the chair Eye-gaze Looking towards the problem Looking everywhere else Facial expression Eyes tightening Eyes widening Raising eyebrows Smile Lowering eyebrow Nose wrinkling Depressing lower lip corner Hand movement Typing, clicking mouse Hands not on the mouse/keyboard
Agents for LMSs w Knowbots (or Knowledge Robots) created to automate the repetitive tasks of human facilitators in online workshops w System developed at ALN Center at Vanderbilt University, Nashville, TN
System architecture w 5 components: knowbots, the learner, the knowledge base, the repository of assignments and the interface with the facilitator. w Knowbots sit between the instructor and the learner, mediating the interaction.
3 types of knowbots : w scheduled - sends a reminder and a report to each participant upon completion of a scheduled check w on-demand - invoked by the learner; these knowbots return results immediately to the requesting user w submission helper - for submission of an assignment that assists the user in submitting the assignment; they also notify the facilitator when the submission is made.
w Knowbot structure: – user-interface agents – checker agents (agents that check submissions) – e-mail agents – knowledge base modules. w User-interface agents - graphical interface, webbased agents; assure user interaction with the knowbot 1. Execute the checker agents by request 2. Present information to the user 3. Provide appropriate interface to execute actions such as requests for help 4. Communicate with other agents and with the knowledge base.
w Email agents are responsible for generating, composing, organizing, and sending e-mails to both the instructor and the participants. w Examples of e-mails that are generated and sent to the participants are: § the assignment-status report § the assignment reminder and notification § the message responding to a request for help. w The e-mail agents compose the content of the e-mail by retrieving data from the knowledge base.
w Checker agents are responsible for checking assignments for the participants. w The agents can be invoked either by the scheduler or by the participant through the user-interface agents. § determine the completion status of the assignment based on the pre-defined knowledge of requirements for assignment completion. § record the results and access the knowledge base through the established Open Database Connectivity (ODBC) using the Cold Fusion Markup Language (CFML). § determine what particular knowledge each participant needs in order to complete the assignment.
Knowbots in the system Posting knowbot - looks for two types of messages posted in the specified forum by participants: one is a self-introduction message, the other is a reply-to-another message. The knowbot then sends a reminder and the results of the scheduled check via e-mail to the participants. S, OD Course Review knowbot - looks for at least 3 course-reviewed messages posted in 3 different threads by the participants and sends a reminder and the result of the checking by e-mail to the participants. S, OD Basic HTML knowbot - checks the status of each participant's personal homepage to determine if it contains the required elements such as mail-to tag, bulleted list, etc. S, OD
Topic knowbot - is invoked by the student and determines if at least one message has been posted into the specified forum in the conferencing system about the required topic. The result is displayed to the student. OD only Multimedia knowbot - Each participant submits information via a knowbot. The knowbot notifies the workshop facilitator about the submission, provides a template for the facilitator to check the participant's work, stores the results into the database and sends a notification e-mail to report the result to the participant. Submission Helper Discussion Builder knowbot - Same functionality as Multimedia knowbot Submission Helper
3. Agents for e-commerce
Electronic commerce w Transactions - business-to-busines (B 2 B) - business-to-consumer (B 2 C) - consumer-to-consumer (C 2 C) Difficulties of e. Commerce w Trust w Privacy and security w Billing w Reliability
Consumer's buying behavior Consumer's Buying Behavior (CBB) research - a number of models of the consumer's behavior CBB - Guttman e. a. , 1998 v Need identification v Product brokering v Merchant brokering v Negotiation v Purchase and delivery v Product service and evaluation - some stages may overlap
Agents as mediators in e. Commerce Persona Logic Need identification Product brokering Merchant brokering Negotiation Purchase and delivery Product service Bargain Firefly Finder Jango Kasbah T@T Intelli. Shoper
(a) Comparison shopping agents Search online shops to find products, merchants and best deals Product brokering Techniques: w feature-based filtering – feature keywords w collaborative filtering – similarities between user’s profiles w constraint-based filtering – specifying constraints (price, date limit)
Persona Logic Product brokering w let the users create preference profiles w allows shoppers to specify constraints on a product and scores the products w CSP engine: hard constraints and soft constraints w 1988 AOL Firefly w helps consumers find products (alert) (Ringo – books, CDs) w ACF = Automated Collaborative Filtering w identifies the shopper's "nearest neighbours" and offers products highly rated by them w 1998 Microsoft
Jango Merchant brokering w finds specifications and product reviews w makes recommendations to the user w submit queries to vendor’s sites and interpret results to identify lowest price items w monitors "what's new" lists, watches for special offers w automates the building of “wrappers” to parse HTML docs and extract product’s features w Web pages are different; exploits: Navigation regularities (easy to find products) Corporate regularities (similar look’n’feel) Vertical separation (use of white spaces) 1999 Excite
(b) Auction bots Agents that can organize and/or participate in online auctions for goods Kasbah Aim = develop a Web-based system in which users can create their own agents to buy and sell goods on their behalf User options: § Create a new buying agent § Create a new selling agent § See currently active agents § Create a new finding agent § Browse the marketplace for active agents
w Selling agent parameters set by the user: - desired date to sell the good - desired price to sell the good - minimum price to sell at - "decay" function of the price over time to determine the current offer price • anxious - linear function • cool headed - quadratic function • frugal - exponential function w Buying agent parameters set by the user - date to buy the item by - desired price - maximum price - "growth" function of price over time
Tête-à-tête w Integrates product brokering, merchant brokering, and negotiation w User agents negotiate across multiple attributes of a transaction, e. g. , warranty length and options, shipping time and cost, service contract, return policy, quantity, accessories, credit options, payment options w Agents quantify those aspects using a multi-attribute utility function w Today: Frictionless Commerce applies the technology to B 2 B markets (e-sourcing)
Intelli. Shoper (U. Iowa) w Integrates product brokering, merchant brokering, and negotiation w Goals: § Customize behavior adaptively by learning user’s preferences § Provides assistance by remaining autonomous from both customer and vendors § Protect shoppers’ privacy by concealing their identities and behavior from vendors
Anonymizing server Intelli. Shoper server Privacy Agent My. SQL Monitor Agent Shopping Persona Learning Agent Vendor plug-ins Vendor Web sites Sequence of shopping assistance activities Basic interaction loop 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. The user creates an account and one or more personae The user takes on a persona The persona initiates a shopping session by submitting a query to the LA The LA stores the user’s request in the database The LA uses vendors plug-ins to send requests to vendors Results from vendors are parsed through the vendors plug-ins IS stores the result in the database The LA uses the persona profile to rank the hits The LA presents the results to the persona The PA forwards the results to the user The user can further interact with the LA
Occurs offline 12. The MA loads standing queries from the database 13. The MA uses vendor plug-ins to check for any new results from the vendors 14. IS parses new and updated hits 15. IS stores the hits in the database until the users logs in again w Privacy Agent w lets the user take a shopping persona w hides identity user info (permutation, stripping of IP addresses, encryption, decription) w Shopping Persona w becomes the “public user” w 2 aims: protect user privacy + multiple profiles w Interface w create new persona + preferred sites w see current personae (name, what to buy, preferences) w submit a new shopping request via the query interface w view hits
w Interface with the vendors Web sites w submitting queries w parsing results w Language for specifying vendor dependent logic – based on XML and inspired by Apple’s Sherlock engine w Persona’s Profile w Preference w Keywords w Relevant features: numeric (discretized) and textual (keywords) w Updating the user profile w Temperatures for features w Updates temperatures after any user action related to a given hit
Temperature update T(t+1) = (1 - ) T(t) + T 5 possible actions: w Buy – string positive feedback T = +2 w Browse – weak positive feedback T = +1 w Skip – weak negative feedback T = -1 w Remove – strong negative feedback T = -2 Status - Research project - Current prototypes: e. Bay, Yahoo and Amazon auctions - Research on the development of intelligent wrappers that could automate submitting queries and parsing results.
References w M. Wooldrige. An Introduction to Multi. Agent Systems, John Wiley&Sons, 2002, Ch. 11, p. 243 -266. w R. Guttman, A. Mokas, P. Maes. Agents as mediators in electronic commerce. In Intelligent Information Agents, M. Klush (Ed. ), Springer Verlag 1999, p. 131 -152. w P. Noriega, C. Sierra. Auctions and multi-agent systems. In Intelligent Information Agents, M. Klush (Ed. ), Springer Verlag 1999, p. 153 -175. w W. Brenner, R. Zarnekov, H. Witting. Intelligent Software Agents, Springer Verlag, 1998, Ch. 6, p. 267 -299. w K. Sycara, Massimo Paolucci, Joseph Giampapa; “The RETSINA MAS Infrastructure”; Tech. Report CMU-RI-TR-01 -05; 2001 w K. Chen, K. Sycaca; “Web. Mate: A Personal Agent for Browsing and Searching”; The Robotics Institute, Carnegie Mellon University; 1998 w K. L. Clarc, V. S. Lazarou; “A Multiagent System for Distributed Information Retrieval on the World Wide Web”; 1997 w F. Menczer, W. Street, A. Monge. Adaptive assistants for customized eshopping. IEEE Intelligent Systems, Nov/Dec 2002, p. 12 -19.
References - continued w S. Case, N. Azarmi, M. Thint, T. Ohtami. Enhancing e-communities with agent-based systems. IEEE Computer, July 2002, p. 64 -69. w R. Ganeshan, W. L. Johnson, E. Shaw, and B. P. Wood. Tutoring Diagnostic Problem Solving , In Proceedings of the Fifth Int'l Conf. on Intelligent Tutoring Systems, 2000. w E. Shaw, W. L. Johnson, and R. Ganeshan. Pedagogical Agents on the Web. In Proceedings of the Third Int'l Conf. on Autonomous Agents, pp. 283 -290, May, 1999. w C. Thaiupathump, J. Bourne, J. O. Campbell. Intelligent Agents for Online Learning. JALN Volume 3, Issue 2 - November 1999. ADELE: http: //www. isi. edu/isd/ADE/ade-body. html w Ganeshan, R. , Johnson, W. L. , Shaw, E. , and Wood, B. P. Tutoring Diagnostic Problem Solving , In Proceedings of the Fifth Int'l Conf. on Intelligent Tutoring Systems, 2000 w Shaw, E. , Ganeshan, R. , Johnson, W. L. , and Millar, D. Building a Case for Agent. Assisted Learning as a Catalyst for Curriculum Reform in Medical Education, In Proceedings of the Int'l Conf. on Artificial Intelligence in Education, July, 1999 w Shaw, E. , Johnson, W. L. , and Ganeshan, R. , Pedagogical Agents on the Web. In Proceedings of the Third Int'l Conf. on Autonomous Agents, pp. 283 -290, May, 1999
References - continued STEVE: http: //www. isi. edu/isd/VET/vet-body. html w Rickel, J. , & Johnson, W. L. , Virtual Humans for Team Training in Virtual Reality, in Proceedings of the Ninth International Conference on AI in Education, pp. 578 -585, July 1999, IOS Press. (Received Best Paper award. ) w Rickel, J. , & Johnson, W. L. , Intelligent Tutoring in Virtual Reality: A Preliminary Report, in Proceedings of the Eighth World Conference on AI in Education, pp. 294301, August 1997, IOS Press. w Rickel, J. , & Johnson, W. L. , Integrating Pedagogical Capabilities in a Virtual Environment Agent, in Proceedings of the First International Conference on Autonomous Agents, pp. 30 -38, February 1997. Survey of Work on Animated Pedagogical Agents w W. L. Johnson, J. W. Rickel, and J. C. Lester. Animated Pedagogical Agents: Face-to. Face Interaction in Interactive Learning Environments. International Journal of Artificial Intelligence in Education 11: 47 -78, 2000. w Johnson, W. L. , Pedagogical Agents, invited paper at the International Conference on Computers in Education. Also to appear in the Italian AI Society Magazine.
Online references Product brokering w Net Perceptions: http: //www. netperceptions. com w CDNOW: http: //www. cdnow. com Merchant Brokering w Priceline: http: //www. priceline. com w Internet Auction List – more than 2500 auction companies listed (as of 2002) http: //www. internetauctionlist. com Auctions bots w Intelli. Shoper: http: //myspiders. biz. uiowa. edu/~fil/intellishoper Software Agents Group, MIT Media Lab http: //agents. media. mit. edu/projects/ w Kasbah - Chaves, Maes, 1996 w Tête-à-Tête - Guttman, Maes, 1998 w Firefly - Shardanand, Maes, 1995 Auction Agents for the Electric Power Industry http: //www. agentbuilder. com/Documentation/EPRI/index. html Information agents w Info. Sleuth: http: //www. argreenhouse. com/Info. Sleuth/index. shtml w Retsina: http: //www-2. cs. cmu. edu/~softagents/retsina_agent_arch. html
Скачать презентацию Intelligent agents on the Web Adina Magda Florea
362ba968d65ca6b0d25ce5a4029ac68e.ppt