Chapter 14 Intelligent Agents Chapter Objectives Explain how intelligent agents are computational entities capable of autonomously achieving goals by executing needed actions. Summarize the general characteristics of intelligent agents. List the different types of intelligent agents. Describe how data mining models, expert systems, and intelligent agents can be collectively applied to find solutions to difficult problems. ____ CD-59 In Chapter 12 we defined an intelligent agent as a computational entity capable of autonomously achieving goals by executing needed actions. In Section 14.1 of this chapter we describe several common characteristics of intelligent agents. In Section 14.2 we categorize intelligent agents by the functions they perform. Finally, in Section 14.3, we describe how intelligent agents, data mining models, and expert systems can be combined to form intelligent systems for solving difficult problems. 14.1 Characteristics of Intelligent Agents Each day we encounter situations that require us to allocate time and energy to tasks we could just as well delegate to someone able to act on our behalf. This individual would know our likes, dislikes, needs, and desires and would be able to learn and modify his or her behavior as our needs change over time.A few of the tasks carried out by such a person might include finding a best price for a product we wish to purchase, notifying us when our favorite author has just published a new book, buying and selling stocks for our stock portfolio, and finding us a date for Saturday night.We can only imagine the salary required to obtain and keep the services of such a useful person! Even if we paid top dollar, few individuals would be able to make this type of commitment on our behalf. Unlike human beings, computer systems are tireless entities that do not complain about performing countless sequences of mundane tasks. Because of this, computer programs are viable candidates for satisfying our wants and wishes. Computer programs able to assist us with everyday tasks and learn from the information we provide are known as intelligent agents. Although there seems to be little general agreement on a precise definition of what constitutes an intelligent agent, four common agent characteristics have been identified (Sycara, 1998): 1. Situatedness. Agents receive sensory information from their surroundings and perform actions based on this information. 2. Autonomy. Agents maintain their own internal state and are able to act without direct intervention from humans or other agents. 3. Adaptivity. Agents are able to react to a changing environment.Agents with this characteristic can learn from experience and create goals to be achieved. 4. Sociability. Agents have the ability to confer with other agents and/or humans. Although few, if any, of today’s intelligent agents possess all of the these characteristics, computer systems exhibiting one or several of these behaviors have been developed. Let’s take a closer look at the types of intelligent agents that do exist. CD-60 Chapter 14 • Intelligent Agents 14.2 Types of Agents Lewis (1998) categorizes three general types of intelligent agents—anticipatory agents, filtering agents, and semiautonomous agents.As the term implies, anticipatory agents attempt to anticipate the intentions of a user.A rudimentary example of an anticipatory agent is Microsoft Word’s animated paperclip.A more complex anticipatory agent might do our bidding on Ebay once we provide a product description, bidding criteria (e.g., only bid on items sold by someone who has sold previous items through Ebay), and a maximum offering price. Lewis further groups anticipatory agents into those able to infer a user’s plan of action and those limited to simple concept learning.A common conceptlearning task is to have an agent use e-mail headers to categorize e-mail messages. Filtering agents are able to carry the categorization process one step further in that they can evaluate, prioritize, and delete information such as incoming e-mail messages. MAXIMS (Lashkari, Metral, and Maes, 1994) and MAGI (Payne and Edwards, 1997) are two systems offering these types of filtering capabilities. Semiautonomous agents deal with tasks requiring complex sequences of actions.The agent is provided with a goal to achieve by interacting with the user through a form-based interface.A typical goal is to have the agent search for an item, such as an airline ticket, which meets a set of constraints. Once found, the agent notifies the user via e-mail.The user is then responsible for making the actual ticket purchase. Haag, Cummings, and McCubbrey (2002) classify intelligent agents into one of four possible categories. Find-and-retrieve agents are semiautonomous agents that surf the internet to locate requested products or information. User agents are anticipatory by nature in that they help individuals perform computer-related tasks. Monitor-and-surveillance agents watch over Web sites, large-scale networks, auction sites, and the like to alert users of changes that may be significant enough to merit immediate action. For example, a business supporting a Web site for selling exercise equipment might have a monitor agent watch for price changes at competing Web sites selling similar or identical products. Finally, data mining agents function within a data warehouse structure to discover changes in business trends. If the agent finds new information of potential interest to the business, management is altered and appropriate actions can be set into place. Case et al. (2001) uses the term proactive to indicate agents able to initiate actions without specific direction from the user. He describes cooperative agents as agents that interact with other agents and perform actions based on the results of the communication. Cooperative agents can be used to build what he describes as e-communities, which allow members to locate other relevant people and stored information. In addition, e-communities support kinship, which makes members feel a sense of belonging to and identifying with a group. Case et al. (2001) also presents a personal agent framework (PAF) developed by Btexact Technologies that integrates several types of intelligent agents. 14.2 • Types of Agents CD-61 Durfee (2001) describes coordination as an agent’s ability to choose its actions in the context of other agents.The purpose of coordination is to achieve some common goal. As an example, the faculty within the management department of a university could each have an agent acting on their behalf during the development of next semester’s class schedule. Individually assigned agents know which courses their associated faculty member prefers to teach, is willing to teach, and does not wish to teach. Likewise, preferred teaching times and room assignments are also made available to each agent. During the scheduling process, individual agents attempt to obtain a best class schedule based on the desires of their associated faculty member.At the same time, each agent will likely find it necessary to compromise some aspects of a best schedule in order to achieve the common goal of developing a semester class schedule. 14.3 Integrating Data Mining, Expert Systems, and Intelligent Agents Intelligent agents, data mining models, and expert systems are similar in that they each use intelligent techniques to solve difficult problems. However, obvious differences exist between the three approaches.With data mining, the emphasis is on applying induction to build models that generalize data. Expert systems also build generalized models, however, the models are constructed by extracting knowledge from one or more human experts. Intelligent agents and data mining models are alike in that each share the ability to learn from their environment. Intelligent agents and data mining models differ in that intelligent agents are goal directed, whereas data mining models are used to test and create new hypotheses about data. Intelligent agents and expert systems have at least three marked differences. First of all, an expert system contains general problem-solving knowledge, whereas the knowledge processed by an intelligent agent is personalized. Second, where an expert system passively responds to questions, an intelligent agent acts on its own accord based on the sensory information it receives from the environment. Finally, where intelligent agents carry out everyday tasks, expert systems perform high-level functions within a specialized domain. As the three approaches differ in their intent, the techniques are complementary and can be used together to build useful models for solving difficult problems. Let’s combine the three approaches to help us build an automated system for performing supervised data mining. Figure 14.1 displays the proposed model. At the center of the model lies a data mining agent.The agent is responsible for coordinating various data mining sessions.A natural language interface allows the user to easily interact with the agent.The user is responsible for presenting the agent with the data to be mined as well as the goals to be achieved.The user can interact with the agent by making suggestions about model choice, model parameter settings, and the like at any time. When presented with a set of data, the agent sends the data to a rule-based domain analyzer for initial analysis.The domain analyzer returns a summary report indicating • CD-62 Chapter 14 Intelligent Agents the location of missing and possibly noisy data.The report also offers summary statistics for all categorical and numeric attributes. After the domain analysis, the data mining agent is responsible for cleaning the data and selecting a set of candidate input attributes. For example, candidate attributes can be selected by applying the genetic attribute selection algorithm described in Chapter 5. In addition, the agent eliminates redundant numeric attributes by computing attribute correlational values. After a set of attributes have been selected, the agent invokes a rule-based model selector, such as the one described in Chapter 12, to choose a data mining technique. Once a technique has been chosen, the agent queries a rule-based parameter selector to obtain parameter settings for the forthcoming data mining session. A sampling of plausible rules for the parameter selector were offered in Section 12.2.The rule-based model and parameter selectors are shown in the top portion of Figure 14.1. With the data mining model and all parameter settings selected, the agent initiates a data mining session.When the session is complete, the agent evaluates and interprets the results by examining test set classification correctness, quality of production rules, as well as any available measures of class quality. If the results satisfy the specified goals, 14.3 • Integrating Data Mining, Expert Systems, and Intelligent Agents CD-63 Figure 14.1 • An agent-based model for data mining Data Mining Session Data Mining Agent Rule-Based Parameter Selector Rule-Based Model Selector Rule-Based Domain Analyzer Natural Language Interface User Summary Report Parameter Settings Cleaned & Transformed Data Data Mining Tool Results Input Data Previous Settings Model Choice Attribute Information New Settings Summary Report Selected Data a summary report is written and the process terminates. If the results are not acceptable, the agent decides on one or more of the following actions: • Invoke the parameter selector to make suggestions for resetting model parameter values • Choose a new data mining model • Modify the selection of input attributes • Conclude that the domain is not amenable for data mining If one or more of the first three selections are chosen, the agent makes the changes described by the choice(s) and initiates a new data mining session.This process continues until an acceptable result is seen or the agent decides to terminate the process. Finally, to support agent learning, the agent keeps a record of each data mining session.When the entire data mining process terminates—either successfully or unsuccessfully—the agent learns by modifying the confidence values for the rules applied during each mining session. Rules that lead to improved results have their confidence factors increased, whereas rules that lead to negative results are given decreased confidence scores. 14.4 Chapter Summary Several types of intelligent agents have been defined; however, all agents ideally share four common characteristics. These characteristics include the ability to receive and act upon information from their environment, the ability to act without direct human intervention, the skill to react to a changing environment, and the talent to confer with other agents and/or humans. Intelligent agents and data mining models are alike in that each share the ability to learn from their environment. However, intelligent agents are goal directed, whereas data mining models are used to test and create new hypotheses about data. Unlike expert systems, designed for high-level problem solving in specialized domains, intelligent agents contain general problem-solving knowledge for performing everyday tasks. As data mining models, expert systems, and intelligent agents differ in their intent, the techniques are complementary and can be used together to build useful models for solving difficult problems. 14.5 Key Terms Anticipatory agent. An intelligent agent that attempts to anticipate the intentions of a user. Cooperative agent. An agent that interacts with other agents and performs actions based on the results of the communication. • CD-64 Chapter 14 Intelligent Agents Coordination. An agent’s ability to choose its actions in the context of other agents. Data mining agent. An intelligent agent that resides inside a data warehouse and attempts to discover interesting business trends. E-community. An electronic community that allows members to locate other people and stored information. Filtering agent. An intelligent agent that is able to categorize and prioritize information. Find-and-retrieve agent. Semiautonomous agents that surf the internet to locate requested products or information. Intelligent agent. Computational entities capable of autonomously achieving goals by executing needed actions. Monitor-and-surveillance agent. An intelligent agent that watches Web sites, networks, and the like and alerts users of significant changes likely to require immediate action. Proactive agent. An intelligent agent that initiates actions without direction from the user. Semiautonomous agent. An intelligent agent that, when provided with a goal to achieve, is able to carry out a sequence of tasks to achieve the goal. User agent. An intelligent agent that helps users with computer-related tasks. 14.6 Exercises Review Questions 1. Several types of agents were described Section 14.2. For each type of agent, decide which one of the four common characteristics of intelligent agents listed in Section 14.1 is the most important agent feature. State your reason for each selection. 2. Does an agent’s ability to coordinate with other agents require agent cooperation? Why or why not? 3. List differences and similarities between the purposes and tasks of data mining models, expert systems, and intelligent agents. 4. Use the find-and-retrieve agent located at www.mysimon.com to search for one or more items. Have the buyer agent at www.bottomdollar.com search for the same item(s). Compare the results obtained by the two buyer agents. 14.6 • Exercises CD-65