6.5 Intelligent Agents
6.5.1 Agent-Based Systems
Due to their specific, machine learning techniques can be successfully employed for building agent-based systems. Such agents would be able to work in partially understood environments used by a large variety of users with a dynamically changed behaviour. These agents are also able to operate in real time, to learn from experience, and to adapt to unforeseen situations. The most important aspect, for the purpose of this thesis consists of the fact that agents are able to capture the unspecified and/or changing preferences of these users, and therefore can serve better the system functionality, for increased usability.
Agent Factory
Several platforms and development tools for agents have been developed, supporting in particular the implementation phase. Some of these platforms comply with the Foun- dation for Intelligent Physical Agents (FIPA) (FIPA, 2003), which is a body currently working in the area of agent standardisation. The most important of these systems are: Zeus (Nwana et al., 1998, 1999), FIPA Open Source (Poslad et al., 2000), JADE — a Java Agent Development Framework (Bellifemine et al., 1999), and LEAP — Lightweight and Extensible Agent Platform (Bergenti and Poggi, 2001).
Agent Factory (O’Hare and Abbas, 1995; O’Hare et al., 2000a; Collier et al., 2003) is an academic prototype, developed and maintained in-house at University College
Dublin. The non-adaptive version of the ECHOES system has been developed on an agent-based architecture, supported by Agent Factory. Thus, this platform has been used for designing the adaptive version of ECHOES VE (see Section 11.4.6).
Agent factory provides a cohesive framework for the development and deployment of agent-based applications. In particular, it provides extensive support for the creation of Belief Desire Intention (BDI) agents. This agent type is realised through the imple- mentation of some mental state architecture and a corresponding agent interpreter that manipulates the mental state, allowing the agent to reason about how best to act.
The framework itself is implemented in Smalltalk-80, while a more recent version has been developed using Java. The Agent Factory runtime environment consists of:
• Agent Virtual Machine which contains the kernel of the Agent Factory platform including the deductive apparatus for the BDI agents;
• Message Transport System which manages the delivery of messages both between agents on the same platform and between agents on different platform;
• Migration Manager which oversees the migration of agents between different plat- forms;
• White Pages Agent which supervises the agent platform and manages the creation and deletion of agents.
BDI concepts are fundamental to Agent Factory and its agent structure reflects this, through the following components describing every agent:
• a mental state,
• commitment rules,
• actuators, and
• preceptors.
An agent’s mental state contains the agent’s current model of itself and its environ- ment. This knowledge is represented as beliefs. Current beliefs refer to beliefs that are held about the current state of the environment at a single point in time, while temporal beliefs have some temporal attribute associated with them. Such beliefs are essential for representing knowledge that is of a persistent nature.
Commitment rules describe the situation under which an agent may adopt a certain commitment. A commitment in Agent Factory is some action that the agent wishes to perform. After deliberating on its belief set and reconciling them with its commitment rules, an agent will adopt a commitment, in terms of committing to carrying out some course of action.
Actuators must be explicitly programmed and their remit is to affect the required changes in the agent’s environment as a result of the commitments it has adapted.
Preceptors constitute the functional units that an agent uses for building a model of its environment. These must be explicitly programmed and generate an appropriate belief set that represents a true model of the current state of the agent’s environment.
6.6
Summary
This chapter provided detailed descriptions of various machine learning algorithms which are used in this thesis. Previously outlined features of these algorithms, fur- ther detailed in Chapter 8, argues for the suitability of this particular set of machine learning techniques, for addressing study objectives. The availability of the software packages supporting the implementation of these techniques constituted an additional factor which guided their choice.
Obviously, other clustering techniques (e.g. K-means clustering (MacQueen, 1967; Anderberg, 1973), hierarchical clustering (Anderberg, 1973), principal components anal- ysis (Jolliffe, 1986; Jain et al., 1999) etc.), rule induction algorithms (McMillan et al., 1992), case-based reasoning algorithms (Riesbeck and Schank, 1989; Aamodt and Plaza, 1994), or neural networks able to deal with time series processing (e.g. recurrent SOM (Koskela et al., 1998), temporal SOM (Kangas, 1990; James and Miikkulainen, 1995) etc.) could have been considered. However, at this point, the interest was primarily to expand the currently used methodologies and tools in order to overcome their lim- itations, rather than proposing the best methodology. Study findings (see Chapter 8) advocate that the proposed machine learning techniques represents a significant method- ological contribution of this thesis, since these algorithms are successfully employed to implicitly capture the knowledge referring to navigational strategies and to cluster users’ trajectories.
This chapter is the last one dedicated to study methodology. The following chapter, the first one among the chapters presenting the study findings, refers to the individual differences related to usability.
Chapter 7
Individual Differences Related to
Usability
Chapter 1 Introduction Background Chapters Chapter 2 Usability of VE Chapter 3 Navigation in VE Chapter 4 Sense of Presence Methodological Chapters Chapter 6Artificial Intelligence for Advanced Data Interpretation and Exploitation Chapter 5 Methodology Chapter 12 Conclusions Results Chapters Chapter 8
Individual Differences in Navigational Patterns Machine Learning Approach
Chapter 9
Individual Differences in Navigational Patterns Geometry of Curve Approach
Chapter 10
Individual Differences in Experiencing Presence Chapter 7
Individual Differences related to Usability
Chapter 11 Towards Accommodating Individual Differences: Design Guidelines
7.1
Introduction
Until now, the thesis has dealt with providing background and a description of the methodology to be adopted. This chapter, presents results of a study on the impact of individual differences on system usability. The study objective addressed within this chapter focuses on the investigation of individual differences related to system usability. This objective is further detailed in:
• the investigation of the individual differences impacting on performance on spatial tasks, and
• the investigation of the individual differences impacting on the level of satisfaction with the system.
These two objectives delineate the two components involved in assessing usability of any artefact: the level of performance of tasks completed through the use of that arte- fact, and the level of satisfaction with interacting with that system (see Section 2.2.1). The performance indicators are usually related toefficiency and effectiveness (ISO, 1997). In this study, efficiency is considered in terms of resources required to perform the task, while effectiveness is related to the accuracy and completeness of performing the task (see Section 2.2.1). Efficiency has been measured in terms of the time needed to complete the search task, while effectiveness has been assessed through the number of collisions encountered during navigation.
User’s satisfaction consists of users’ attitude regarding their interaction with the system, and it was measured with a self-rating questionnaire (see Section 5.6.1).
This study design takes the shape of a quasi-experiment, where the quasi indepen- dent variables (IVs) can only be statistically, rather than directly manipulated. Thus, participants are assigned to a particular condition because they already qualify for that condition, such as males or females; experts or novices (see Section 5.6). The quasi IVs are users’ gender, their previous computer games experience, their personality cognitive style, and the level of sense of presence induced during interaction. By statistical ma- nipulation, it was aimed to investigate the effect of these variables on system usability expressed in terms of task performance and level of satisfaction. These two latter as- pects represent the dependent variables (DVs). Several working hypotheses have been formulated, whose theoretical support has been discussed in Chapter 4:
• H1. Males achieve better task performances than females.
• H2. Males experience a higher level of satisfaction in rapport with the VE system.
• H3. The greater the users’ computer games experience, the better the task per- formances.
• H4. The greater the users’ computer games experience, the higher the satisfaction in rapport with system usability.
• H6. Different dimensions of personality cognitive style impact on task perfor- mance.
• H7. The greater the presence experienced by the users, the higher their satisfac- tion.
This chapter focuses on testing these hypotheses, in the light of the objectives out- lined above. The description of study results is followed by their discussion.