intelligent_agents by xuyuzhu


									Intelligent Agents

    With Java

                                    Focus of talk
• A basic look at agent-based reasoning, modeling,
  and learning
• How agents can enhance the capability and
  productivity of commercial application software
• The effect of agents on the Web, with a Java twist

         Artificial Intelligence: Introduction
• The science of AI is approximately forty years
   – dating back to a conference at Dartmouth in 1958
• The public perception of AI has not always
  matched the reality
• The excitement of both scientists and the popular
  press tended to overstate the real-world progress
  of artificial intelligent systems
• Early success promised rapid progress towards
  practical machines intelligence. Areas of early
   – Game playing, mathematical theorem proving, common-
     sense reasoning, college mathematics

                                   Introduction, contd.
• AI research labs began specializing in narrow
   – Speech recognition
   – Natural language understanding
   – Image optical character recognition
• The early successes were followed by a slow
  realization that things that humans do with very
  little effort was near impossible for the computer
   – What was hard for people and easy for the computer was
     more than offset by the things that were easy for people to
     do but almost impossible for computers to do

                                 Introduction, contd.
• The promise of the early years has never been
  fully realized
• The term artificial intelligence have become
  associated with failure and over-hyped technology
• Nevertheless, researchers in AI have made
  significant contributions to computer science
   – WIMP (Windows, icon, mouse, pointer) user interface
      • Considered highly controversial and impractical when
        first introduce by the IA community
   – Object-oriented programming techniques
      • Refinement of the AI Frames concept

                                         Basic Concepts
• AI has always focused on problems which lie just
  beyond the reach of state-of-the-art computers
   – Effectively pushing the current bleeding-edge technologies
• As computer science and computer systems
  evolved, the focus and areas which falls into AI
  research have also changed
• We can identify three major phases of
  development in AI research

                                                First Phase
• Much of this work dealt with formal problems
  that were structured and had well-defined problem
   – Math related skills: proving theorems, geometry, calculus,
     games (checkers, chess)
• Emphasis was on creating general “thinking
  machines” capable of solving broad classes of
   – These systems tended to include sophisticated capabilities
     relating to reasoning and search techniques

                                              Second Phase
• Marked by the recognition that the most successful AI
  projects were aimed at very narrow problem domains
   – These systems usually encoded much specific knowledge about the
     problem to be solved
   – This approach of adding specific domain knowledge to a more
     general reasoning system led to the commercial success in AI –
     Expert Systems.
• Rule-based expert systems were developed to do many
   – Chemical analysis, configuring computer systems, diagnosing
     medical conditions in patients
   – Suitable for repetitive and hazardous work
   – Automated Process Control (Manufacturing Systems)

                                 Second Phase, contd.
• Expert systems utilized research in a number of AI
  based discipline
   – Knowledge representation, knowledge engineering,
     advanced reasoning techniques
   – These systems proved that artificial intelligence could
     provide real value in commercial applications
• Expert systems workstations with powerful
  integrated development environments were
   – Lisp, Prolog, Smalltalk
   – These were years ahead of commercial software

                                             Third Phase
• Since the late 1980s much of the AI community
  has been working on solving some difficult
   – Machine vision and speech
   – Natural language understanding and translation
   – Commonsense reasoning and robot control
• Connectionism regained popularity and expanded
  the range of commercial applications through the
  use of neural networks for use in
   – Data mining
   – Modeling
   – Adaptive control

                           Third Phase, contd.
• The AI playing field has been reenergized by
  biological methods such as genetic algorithms and
  alternative logic systems such as fuzzy logic
• Recent explosive growth in the Internet and
  distributed computing has led to the idea of
  Software Agents
• Software Agents are autonomous entities that
  move through the network, interacting with each
  other and performing tasks for their users

                           Intelligent Agents
Intelligent agents are software agents that use
the latest AI techniques to provide
autonomous, intelligent, and mobile software
components, thereby extending the reach of
users across networks

                                                   Foot Note
• Using commercial success as a measure of the
  value of technology is problematic to say the least
• I hypothesize that technology that is most
  beneficial to humanity on a whole will be the least
  commercially viable
   – The rules of supply and demand will not apply to
     technologies that have the following characteristics
       • Source is abundant (water for instance)
       • The ability to transform and make readily available is
         attainable by every society
           – Low technological barrier

            What do we mean by intelligence?
• Do we mean that our agents acts like a human?
  Think like a human? That it acts or thinks
   – There are as many answers as there are researchers involved
     in AI work
• From a software development perspective an
  intelligent agent is one that acts rationally
  primarily from a behavioral view point
   – It does the things we do, but not necessarily the same way
     we would do them
   – Our agent may not pass the Turing test as a yardstick for
     judging computer intelligence

                                          Why AI Failed
• This is only my opinion
   – AI as we know it lacks a true model of cognition that can
     shed insights into events such as
       • Correlation of facts, inference, and memory
       • How the human brain work: higher level cognitive
         functions such as reasoning
   – The Von Neumann model of a computer is a not a
     reasonable model of the brain and of human cognition

          What do we mean by intelligence?
• Our agents will perform useful tasks for us
• They will make us more productive
• They will allow us to do more work in less time,
  and see more interesting information and less
  useless data
• Our programs will be qualitatively better using AI
  techniques than they would be otherwise
• The humble goal of intelligent agents is to develop
  better smatter applications

                               Areas to Explore
•   Symbol processing
•   Neural networks
•   The Internet and the World Wide Web
•   Events-Conditions-Actions

Intelligent Agents


                                  Intelligent Behavior
• There are many behaviors to which we ascribe
   – The ability to recognize situations or cases is a type of
       • For example, a doctor who talks with a patient and
         collects information regarding the patient’s symptoms
       • Then able to accurately diagnose an ailment and the
         proper course of treatment
   – The ability to learn from a few examples and then generalize
     and apply that knowledge to new situations is another form
     of intelligence
• Intelligent behavior can be produced by the
  manipulation of symbols

                                   Symbol Processing
• Symbol Processing is an AI technique
• Assertion: Intelligent behavior can be produced by
  the manipulation of symbols
   – A primary tenets of AI techniques
   – Symbols are tokens which represents real-world objects or
• In this approach, a problem must be represented
  by a collection of symbols
   – An appropriate algorithm must then be developed to process
     these symbols

                           Symbol Processing, contd.
• Physical symbol systems hypothesis
   – Newell and Simon 1980
   – States that only a “physical symbol system has the necessary and
     sufficient means for general intelligent action.”
   – Basic thesis is that intelligence flows from the active
     manipulation of symbols
      • This was the cornerstone on which much of the subsequent
         AI research was built
• Research built intelligent systems using symbols
   – pattern recognition, reasoning, learning, planning
• History has shown that symbols may be appropriate
  for reasoning and planning
   – Pattern recognition and learning are suited for other approaches

                          Manipulation of Symbols
• Symbols in the formulations of If-Then rules
   – Processed using forward and backward chaining reasoning
   – Forward chaining: system deduce new information from a
     given set of input data
   – Backward chaining: system reach conclusion based on a
     specific goal state
• Semantic Network
   – Symbols and the concept they represent are connected by
     links into a network of knowledge that can then be used to
     determine new relationships
• Frames – similar to Objects in the OO paradigm
   – Attributes of a concept are grouped together with related
     procedures for processing
           Symbol Processing and Cognition
• Symbol processing
  – These techniques represent a relatively high level in the
    cognitive process
  – Correspond to conscious thought, where knowledge is
    explicitly represented, and the knowledge itself can be
    examined and manipulated
• Symbol-less approach
  – An approach that is modeled after the brain

                                       Neural Networks
• This technique defines the connectionism camp of
  artificial intelligence
   – More focus on how human or natural intelligence occurs
   – Humans have neural networks, consisting of hundreds of
     billions of brain cells called neurons
   – Neurons are connected by adaptive synapses which act as
     switching systems between the neurons
• Artificial neural networks
   – These are based on the massively parallel architecture found
     in the brain
   – They process information by processing large amounts of
     raw data in a parallel manner

Neuron       Switching
         (Adaptive Synapses)   Neuron


                           Neural Networks, contd.
• Operations of neural networks
   – Different formulations of neural networks are used to
      • Segment or cluster data, classify data, make predictive
        models using data
   – A collection of processing units which mimic the basic
     operations of real neurons is used to perform these functions
• Learning or training
   – As the neural network learns or is trained, a set of
     connection weights between the processing units is modified
     based on the perceived relationship in the data

                      Learning in Neural Networks

                                                Processing Unit
                     Connection Weight          (Collection of
Processing Unit                                    Neurons)
(Collection of

                              Processing Unit
Connection Weight             (Collection of

   Processing Unit                                          Processing Unit
   (Collection of                                           (Collection of
      Neurons)                                                 Neurons)
                              Connection Weight

 Neural Network and Cognitive Functions
• Neural networks
   – Compared to symbol processing systems, neural networks
     perform relatively low-level cognitive functions
   – Knowledge gain through learning is stored in the connection
     weights and is not available for examination & manipulation
• Adaptability
   – The ability of neural networks to learn from and adapt to
     their surrounds is a crucial function needed by intelligent
     software systems
• Cognition
   – From a cognitive science perspective, neural networks are
     more like the underlying pattern recognition and sensory
     processing that is performed by the unconscious levels of the
     human mind

                      The Internet and the WWW
• The Internet grew out of government funding for
  high energy physics researchers who needed to
  collaborate over great distances
• Byproduct of solving the communication problem
   – Developed protocols that allows different computers to talk
     to each other, exchange data, and work together
   – TCP/IP became the de facto standard networking protocol
     for the Internet
• Astounding Growth in the Internet
   – Exponential growth in the number of sites
   – Thousands of new sites are connected to the Internet each

          The Internet and the WWW, contd.
• Internet Services
   – Electronic mail was once the primary service provided by
     the Net
   – Information publishing and software distribution are now of
     equal importance
       • The Gopher text information service: early 1990s
       • Gopher was the information publishing on the Net
   – FTP provides valuable services
       • Download research papers and articles, retrieve software
         updates, and download complete software products
   – It was HTTP that brought the Internet from the realm of
     academia and computer technologists into the public

                     The Internet and the WWW
• Mosaic browser: University of Illinois
   – Transformed the Internet into a general-purpose
     communication medium
   – Computer novices and experts, consumers, and businesses
     interact in entirely new ways
   – The Net has become a new business platform
• Web Services
   – The Web publishing and broadcasting capabilities has
     extended the range of applications and services
       • VoD, streaming audio and video, etc
   – The ubiquitous Web browser provides a universal interface
     to applications regardless of server platform
   – In the browsing or “pull” mode, the Web allows individual
     to explore vast amounts of data in a seamless environment
                                             Web Services
• Limitations of the Browsing or Pull model
   – The basic problem is that knowing that all the information is
     out there but not knowing exactly how to find it
   – This can make the Web browsing experience quite
• Search engines
   – Search engines and Web index sites such as Alta Vista,
     Excite, Yahoo, and Lycos provide important services by
     grouping information by topics and keywords
   – Web browsing is still a hit or miss proposition (with misses
     more likely than hits)

                                 Web Services, contd.
• Intelligent Agents
   – In the current Web environment, intelligent agents will
     emerge as truly useful personal assistants
   – Perform tasks such as searching, finding, and filtering
     information from the Web, and bringing it to a users
• The Evolving Web
   – The Web is evolving into a “push” or broadcast mode, where
     users subscribe to sites which send out constant updates to
     their Web pages
   – In the broadcast mode, the requirement for filtering
     information will not go away
       • Unless the broadcast sites are able to send out
         personalized streams of information

Intelligent Agent


                  From AI to Intelligent Agents
• Whenever a technical field provokes commercial
  interest, this normally results in intense inertia
  towards market positioning
• AI and Commercial Interest
   – The same is true for the AI community
   – There has been a large movement and change of focus in the
     AI research community to apply the basic artificial
     intelligence techniques to a host of commercial interest
       • Distributed computer systems, company wide intranets,
          the Internet, and the WWW
   – Early focus was on word searches, information retrieval, and
     filtering tasks

       From AI to Intelligent Agents, contd.
• Intelligent Agents and Commercial Interest
   – Web in evolving into a collaborative commerce (c-
     commerce) environment – transactions are becoming
     increasing distributed in nature
   – There significant interest in having smart agents which can
     perform specific actions
   – Many researchers have turned their focus to looking at how
     intelligent agents could cooperate to achieve tasks on
     distributed computer systems
       • There is finally a problem in search of a technology
       • As opposed to the other way around
   – Intelligent Agents can provide real value to users in this
     new, interconnected, and networked world

• Abstract look at software agents
   – We have discussed artificial intelligence and its evolution
     into software agents at an abstract level
• We will now take a brief tour of
   – The technical facets of intelligent agents
   – How they work
   – How we classify them based on their abilities and underlying

• Scenario
  – Suppose we have an intelligent agent, running
    autonomously, primed with knowledge about the tasks we
    required of it.
  – The agent is ready to move out on the network when the
    opportunity arises.
• Now what?
  – How does the agent know that we want it to do something
    for us, or that it should respond to someone who is trying to
    contact us?
  – This is where we have to deal with events, recognize
    conditions, and take actions



If (event1,event2=condition)
• Events
  – An event is anything that happens to change the environment
    of which the agent should be aware
     • Arrival of a new piece of mail, change to a Web page, a
       timer going off at mid-night
  – Would like to have asynchronous notification of events
     • Agent would not have to be engaged in busy wait or
       polling for events
     • Agents can sleep, think about what has happened during
       the day, do house keeping tasks, etc, while waiting for
       the next event
  – Event notification
     • When an event occur, the agent has to recognize and
       evaluate what the event means an then respond to it

• Condition/Recognition
  – Determining what the condition or state of the world is,
    could be simple or extremely complex depending on the
  – New mail is a self-describing event
      • The agent may then query the mail system to find out
        who sent the mail, what the subject is, or scan the mail
        text for keywords
      • All of this is part of the recognition phase
  – The initial event may wake up the agent, but the agent has to
    determine the significance of the event in terms of its duties

• Condition/Recognition/Action
  – If intelligent Agents are going to make our lives easier or
    more interesting, they must be able to take action, to do
    things for us
  – Having an agent take an action for us requires a certain leap
    of fait or at least some level of trust
      • We must trust that our intelligent agent is going to
        behave rationally and in our best interest
      • Like all situations where we delegate responsibility, we
        have to weigh the risks and rewards
      • Risk: agent could mess things up, more work to get it
      • Reward: we are free from performing that piece of work

                                Processing Strategies
• Reactive or reflex agents
   – These are one on the simplest types of agents. They respond
     in the event-condition-action mode
   – Reflex agents do not have internal models of the world
   – They respond solely to external stimuli and the information
     available from their sensing of the environment
   – Like neural networks, reactive agents exhibit emergent
     behavior – interactions of simple individual agents
   – Reactive agents share low-level data when they interact, not
     high-level symbolic knowledge
   – Reactive agents are grounded in physical sensor data and not
     at the artificial symbolic space
   – Applications of reactive agents have been limited to robots
     which use sensors to perceive the world

                                Processing Strategies
• Deliberative or goal-directed agents
   – These agents have domain knowledge and the planning
     capability necessary to take a sequence of actions in the hope
     of reaching or achieving a specific goal
   – Deliberative agents may proactively cooperate with other
     agents to achieve a task
   – They may use any and all of the symbolic artificial
     intelligence techniques which have been developed over the
     past forty years

                               Processing Strategies
• Collaborative agents
   – These agents work together to solve problems
   – Communication between agents is an extremely important
       • Each individual agent is autonomous
       • The synergy resulting from their cooperation makes them
         interesting and useful
   – These agents can solve large problems which are beyond the
     scope of any single agent and they allow a modular approach
     based on specialization of agent functions or domain
       • Complex engineering projects – verify different aspects
         of the design.


To top