Application Areas � Agents are usefully applied in domains where LECTURE 11: autonomous action is required. � Intelligent agents are usefully applied in domains Applications where flexible autonomous action is required. This is not an unusual requirement! Agent technology gives us a way to build systems that mainstream software engineering regards as hard! An Introduction to MultiAgent Systems � Main application areas: http://www.csc.liv.ac.uk/~mjw/pubs/imas � distributed/concurrent systems � networks � human-computer interfaces 11-1 11-2 Domain 1: Distributed Systems Domain 2: Networks � In this area, the idea of an agent is seen as a � There is currently a lot of interest in mobile natural metaphor, and a development of the agents, that can move themselves around a idea of concurrent object programming. network (e.g., the Internet) operating on a user’s behalf � Example domains: � This kind of functionality is achieved in the � air traffic control (Sydney airport) TELESCRIPT language developed by � business process management General Magic for remote programming � power systems management � Applications include: � distributed sensing � factory process control � hand-held PDAs with limited bandwidth � information gathering 11-3 11-4 Domain 3: HCI Agents on the Internet � The potential of the internet is exciting � One area of much current interest is the use of agent in interfaces � The reality is often disappointing: � The idea is to move away from the direct manipulation � the Internet is enormous — it is not paradigm that has dominated for so long always easy to find the right information � Agents sit ‘over’ applications, watching, learning, and manually (or even with the help of search eventually doing things without being told — taking the engines) initiative � Pioneering work at MIT Media Lab (Pattie Maes): � news reader � web browsers � mail readers 11-5 11-6 1
Agents on the Internet Agents on the Internet � What we want is a kind of ‘secretary’: someone who � systematic searches are difficult: understood the things we were interested in, (and human factors : we get bored by slow response � the things we are not interested in), who can act as times, find it difficult to read the WWW rigorously (it ‘proxy’, hiding information that we are not interested is designed to prevent this!) get tired, miss things easily, misunderstand, and get sidetracked in, and bringing to our attention information that is of interest organizational factors : structure on the net is only � superficial — there are no standards for home � This is where agents come in! pages, no semantic markup to tell you what a page � We cannot afford human agents to do these kinds of contains tasks (and in any case, humans get suffer from the � the amount of information presented to us drawbacks we mentioned above) leads to ‘information overload’ � So we write a program to do these tasks: this program is what we call an agent 11-7 11-8 Another Scenario A Scenario � ‘The ‘agent’ answers the phone, recognizes the callers, disturbs � Here is a scenario illustrating the kinds of properties you when appropriate, and may even tell a white lie on your behalf. that we hope Internet agents will have: The same agent is well trained in timing, versed in finding Upon logging in to your computer, you are opportune moments, and respectful of idiosyncrasies. ’ (p. 150) presented with a list of email messages, sorted into ‘If you have somebody who knows you well and shares much of order of importance by your personal digital your information, that person can act on your behalf very assistant (PDA). You are then presented with a effectively. If your secretary falls ill, it would make no difference if similar list of news articles; the assistant draws your the temping agency could send you Albert Einstein. This issue is attention to one particular article, which describes not about IQ. It is shared knowledge and the practice of using it in hitherto unknown work that is very close to your your best interests.’ (p. 151) own. After an electronic discussion with a number of ‘Like an army commander sending a scout ahead . . . you will other PDAs, your PDA has already obtained a dispatch agents to collect information on your behalf. Agents will dispatch agents. The process multiplies. But [this process] started relevant technical report for you from an FTP site, in at the interface where you delegated your desires.’ (p. 158) the anticipation that it will be of interest. (From Being Digita l, by Nicholas Negroponte, Hodder & Staughton, � Demonstrator systems used today 1995.) 11-9 11-10 Email Reading Assistants Email Reading Assistants � The staple diet of software agent researchers… � Situation characterized by features of event: � Pattie Maes developed MAXIMS – best known email assistant: � sender of email ‘learns to prioritize, delete, forward, sort, and � recipients archive mail messages on behalf of a user … ’ � subject line � etc. � MAXIMS works by ‘looking over the shoulder’ of a user, and learning about how they deal with � When new situation occurs, MAXIMS email matches it against previously recorded rules � Each time a new event occurs (e.g., email � Tries to predict what the user will do — arrives), MAXIMS records the situation → generates a confidence level action pairs generated 11-11 11-12 2
Agents for E-Commerce Email Reading Assistants � Confidence level matched against two thresholds: � Another important rationale for internet “tell me” and “do it” agents is the potential for electronic Confidence < “tell me”: commerce – agent gets feedback � Most commerce is currently done “tell me” < confidence < “do it”: – agent makes suggestion manually . But there is no reason to Confidence > “do it”: suppose that certain forms of commerce – agent acts could not be safely delegated to agents. � Rules can be “hard coded”; even get help from other � A simple example: finding the cheapest users copy of Office 97 from online stores � MAXIMS has a simple ‘personality’, (a face icon), communicating its ‘mental state’ to the user 11-13 11-14 Agents for E-Commerce Agents for E-Commerce � More complex example: flight from Jango (Doorenbos et al, Agents 97) is � Manchester to Dusseldorf with veggie good example of e-commerce agent meal, window seat, and does not use a Long-term goals: � fly-by-wire control Help user decide what to buy 1. � Simple examples first-generation e- Finding specs and reviews of products 2. commerce agents: Make recommendations 3. � BargainFinder from Andersen Comparison shopping for best buy 4. � Jango from NETBOT (now EXCITE) Monitoring “what’s new” lists 5. � Second-generation : negotiation, Watching for special offers & discounts 6. brokering, … market systems 11-15 11-16 Agents for E-Commerce Agents for E-Commerce � Isn’t comparison shopping impossible? WWW pages all different! � Two key components of Jango/ShopBot: � Jango/ShopBot exploits several regularities � learning vendor descriptions in merchant WWW sites: � comparison shopping � navigation regularity : sites designed so that products easy to find � corporate regularity : sites designed so that pages have same look’n’feel � vertical separation : merchants use whitespace to separate products 11-17 11-18 3
� FAQ-finders: Real Soon Now The idea here is to direct users to FAQ documents in � (Etzioni & Weld, 1995) identify the following specific types of order to answer specific questions. Since FAQS tend agent that are likely to appear soon: to be knowledge intensive, structured documents, � Tour guides: there is a lot of potential for automated FAQ servers. The idea here is to have agents that help to answer the � Expertise finders: question ‘where do I go next’ when browsing the WWW. Such agents can learn about the user’s preferences in the same Suppose I want to know about people interested in way that MAXIMS does, and rather than just providing a temporal belief logics. Current WWW search tools single, uniform type of hyperlink actually indicate the likely would simply take the 3 words ‘temporal’, ‘belief’, interest of a link. ‘logic’, and search on them. This is not ideal: LYCOS � Indexing agents: has no model of what you mean by this search, or Indexing agents will provide an extra layer of abstraction on what you really wan t. Expertise finders ‘try to top of the services provided by search/indexing agents such understand the users wants and the contents of as LYCOS and InfoSeek. The idea is to use the raw information provided by such engines, together with information services’, in order to provide a better knowledge of the users goals, preferences, etc., to provide a information provision service. personalized service. 11-19 11-20 4
Recommend
More recommend
