16

In Chapter 6 we propose that a user's willingness to delegate an activity to an

agent also will depend on the risks involved. Konstan et al. (1997) discuss the

consequences of errors in recommendation agents. If an agent is making a

recommendation, a sample item can be desirable or undesirable, and the agent can

predict it to be good or bad. When the agent predicts a desirable item to be good, the

result is a hit. Predicting an undesirable item to be bad is a correct rejection. Errors occur

when an undesirable item is predicted to be good (a false positive), or when a desirable

item is predicted to be bad (a miss). The consequences of each type of error will vary

with the domain. A recommendation for a restaurant that turns out to be undesirable (a

false positive) is costly in time and money, whereas the time that it takes for the user to

reject a recommendation of an undesirable research article (another false positive) is

minimal. Similarly, missing a legal citation could have serious consequences, whereas

missing a recommendation for a good music CD is less important (Konstan et al. 1997).

1.4.3 Decision Support

Simon's (1960) three-stage model of decision-making provides a starting point

for our discussion of decision support. Simon proposed that decision-making activities

fall into one of three stages:

1. Intelligence - Scanning the environment for conditions that require a decision

2. Design - Identifying, developing and analyzing alternative courses of action

3. Choice - Selecting a particular course of action from those available

A decision-maker rarely follows through these stages in a linear fashion. Many

decisions are broken down into smaller or staged decisions that require circling through
 17

the stages until all of the sub-decisions have been made. Subdivision of the problem can

take place at many levels and at each level decision situations may require that the

decision-maker backtrack to earlier stages in the process before finally reaching a

decision.

Information systems and DSS systems have been created to support users in all

three stages of decision-making (Turban and Aronson 1998). In Chapter 3 we introduce

the idea of search support agents and choice support agents. Search support agents work

in the design phase of Simon's model They help the user identify and find relevant

information about alternatives. Choice support agents may work in both the design and

choice stages. They can help the user analyze alternatives (design), make

recommendations on these alternatives (design) or choose a course of action

autonomously (choice).

In the economic or rational theory of decision-making the decision-maker is

assumed to have well-defined preferences, is able to assign a utility value to each

alternative and chooses the alternative that optimizes utility (Bettman et at 1998). The

information processing theory of decision making (Simon 1955) proposes that people

have limited cognitive capabilities, in both working memory and their ability to process

information. As a result they operate under what Simon called "bounded rationality". To

reduce cognitive effort, people rely on heuristics and are often satisfied with a decision

where the expected outcome is "good enough", rather than optimal. This is called

"satisficing" behaviour.
 18

Behavioural decision-making theory has evolved from Simon's theories and

other observations that people do not always make economic decisions. A number of

systematic deviations from optimal and rational behaviour have been shown to persist in

human decision-making and behavioural scientists have tried to understand how the use

of heuristics and processing strategies can account for these deviations. As well as

cognitive factors, they have also addressed social and psychological factors that affect

decision-making (Elam et ale 1992).

The study of consumer decision-making behaviour applies the behavioural

science approach to a specific domain. Researchers have looked at how consumers

search for information (from both external sources and through memory) as well as how

they select and use choice strategies (Elam et ale 1992).

Many factors can affect the choice of decision strategies. Research has shown

that consumers often do not start a decision task with known preferences - they

"construct" them when required. Decision strategies can be characterized according to

whether information is processed consistently or selectively (how much of the available

information is used), whether information is processed by alternative or by attribute (the

pattern of processing), and whether the strategy is compensatory (requiring trade-offs) or

non-compensatory (not requiring trade-off) (Bettman et al. 1998).

It has long been the goal of Information Systems (IS) developers to design

systems that can assist in overcoming the cognitive limitations of users (Davis and

Olson 1985). More specifically, the study of DSS examines how information systems

can help decision-makers make better decisions.

 19

"A DSS is a computer-based system used by managers as an aid to decision-

making in semi-structured decision tasks through direct interaction with data and

models" (Benbasat and Nault 1990, p. 203). While DSS were originally developed and

implemented for managerial use in organizations, the Internet and the WWW now

provide possibilities for DSS to be used by more individuals and for a variety of

business and personal decisions. Retirement planning is one example of an application

\

where individuals can use the WWW to access various models for forecasting their

needs and develop an appropriate savings plan. DSS researchers are looking at ways to

make these systems more accessible, including the use of agents to search for and find

appropriate systems or components (e.g., Lang and Whinston 1999; Gregg et aI. 2002).

In Chapter 3 we elaborate on how current research in DSS is incorporating

artificial intelligence techniques to add structure to larger and more complex areas of the

decision-making process.

1.5 Summary

This chapter has introduced the various projects that combine in this dissertation.

A common theme throughout these projects is that software agents are viewed as an

innovation. We want to identify areas where agent applications are most likely to be

adopted and how to design agent applications that people will choose to use. The

research questions we pose are directed at understanding where we can best direct

research and development efforts.

 20

Our research uses the idea of decision support to unite the research areas of

software agents and electronic commerce. We have oriented the reader by providing

starting points in each of these top-level areas and provided brief descriptions of some

general directions our research follows through these areas.

 Chapter 2

Approaches to Software Agents - A User's Perspective

~-:;-~\
, Agents' Support) Commerce)

:~~ ~ Cnapterz"' ~" '=

AppIoaches to A~ms "~

! l)efin~ioo t:A Concept

-ClaG ... iieuboo :!.dl¢roiil

Figure 2-1 - Chapter 2 in the Research Overview

This chapter presents an original classification system for agent applications.

This system emerged from a broad review of the literature on software agents. Its

purpose is to communicate to the reader the types of agents that the rest of this research

will address as well as those that are not being addressed.

Under this system. agents are classified in very general terms according to how

the user views and interacts with the agent. This perspective is consistent with a product-

focused approach to agent design and development.

21
 22

2.1 Background

Prior to the mid-1990s, the concept of software agents had been limited to the AI

research community. In July 1994, a special issue of the Communications of the ACM

sparked broadened interest from other disciplines and attention from the popular

computing press (Nwana and Ndumu 1999; Wooldridge and Decker 2000).

In Chapter 1 we introduced two areas where agent technologies are considered to

be important. The human-computer interface is one area where "fIrst generation"

literature promoted the usefulness of agent technology (Negroponte 1995; Bradshaw

1997). As information technology becomes increasingly integrated into everyday tasks,

the need to "delegate, not manipulate" (Negroponte 1995) may be required to exploit the

full potential of current and future systems. Today's human-computer interface is often

considered to be a bottleneck for both experienced and naive users.

After 1994, many "fust generation" books and articles on the subject also

proposed that software agent technologies provide new paradigms for abstraction and

modularity when designing applications for complex, open and distributed systems

(Jennings and Wooldridge 1998) (Bradshaw 1997). The growth and commercialization

of the world WWWhas continued to provide challenges and real-life problems for agent

researchers and developers to address (Nwana and Ndumu 1999; Wooldridge and

Decker 2000). The convergence of computing and communications systems also has

introduced some unique distributed design challenges associated with mobile

computing.
 23

From the introduction to software agents in Chapter 1 we recall that there is still

no commonly accepted defmition for what constitutes a software agent. The differences

between an interface focus and a distributed design focus is signaled in definitions

proposed by researchers working in each stream. Focusing on the interface, Bradshaw

(1997) suggests that agents can be defmed "by attribution", when the user believes that

they are delegating a task to an agent. On the other hand, distributed design research

defines an agent as "a computer system that is situated in some environment and that is

capable of autonomous action in this environment in order to meet its design objectives"

(Wooldridge 1999 p. 29). The user does not playa role in this definition.

In the following section (2.2) we develop a description of three approaches to

agent applications from a user's perspective, where some of the fundamental differences

between these two streams are illuminated. Rather than dismissing agent applications

that do not meet the definition adopted by anyone approach, a more inclusive attitude

can encourage development activity and cross-pollination of research between

approaches.

In section 2.3 we describe how this classification system was applied to a small

sample of research articles about agents. In section 2.4 we submit the classification

schema to a reliability test and apply it across a broader sample of articles from different

fields of research.
 24

2.2 Classification by Approach

A broad review ofthe existing literature on agents was begun in 1998 as part of

this researcher's preparations for this dissertation. It was soon discovered that none of

the existing classifications and taxonomies of agents were effective in organizing this

extensive collection of research or communicating the types of agent applications that

we wanted to include and exclude in our research.

The literature described some agent applications that were completely hidden

from the user. lfthe user does not know he or she is using an agent, we cannot treat it as

an innovative product to·be considered for adoption from the user's point of view.

Another group of agents, often called "interface agents", presented avatars or

cartoon characters that employed natural language processing (NLP) capabilities as a

way of encouraging a more natural and social way of communicating with systems.

These types of agents pose similar research questions as far as the adoption of

innovation, but they also have their own set of research problems. How do users react to

the introduction of anthropomorphism to the interface? What are the social and

emotional dimensions of the interaction between this type of agent and the user? These

are important questions, but they deserve a separate inquiry.

Over a two-year period we developed a "personal" classification system

consisting of three "approaches" to agents: the AI approach, the computer

science/systems (CS/S) approach and the DSS approach. In this section we describe

these approaches and how we refined the classification system by trying to apply it to a

sample of articles on agents.

 25

2.2.1 The AI Approach

Traditionally, AI attempts to model and replicate human intellectual activity and

AI systems are designed as black box systems that focus on results. The classic "Turing

test", where an AI system is expected to produce behaviour that is indistinguishable

from that of a human being, is evidence of this focus (Turing 1950). There is a

significant body of agent research and development that attempts to simulate human

behaviour in a similar way. The defining characteristic of the AI approach to agents is an

objective of producing "realistic" or "believable" behaviour. 1

One type of agent that follows the AI approach is where the end user interacts

with the system through an agent, as if it were a human agent. These agents are often

called "Interface Agents." Improvements to the current generation of direct

nianipulation interfaces are desired to make it easier for users to make full use of the

many features being added to applications. The expanding functionality of the WWW is

also drawing more and more inexperienced users to the desktop computer and simplified

interfaces may help these users. Thus, the "delegate, not manipulate" (Negroponte 1995)

proposition for the next generation of interfaces is an important one.

The original "strong hypothesis" for interface agents proposed that the agent

would be able to observe and learn from the user's actions, providing personalized,

intelligent assistance and tutoring (Nwana and Ndumu 1999). An example would be

I By labeling one of our classifications as the "AI approach" our intention is not
to diminish the contributions of AI in other approaches. Reasoning and learning
capabilities developed within AI provide the autonomous and adaptive behaviour
required for all agent applications.