Relatively few relational agents have been developed and the majority of these can be found in applications related to health and behavioural change. One such example is Laura, the first relational agent (2005). Implemented as part of the MIT FitTrack system she was a computer generated exercise advisor to whom users reported the extent of their exercise [9-14]. Laura’s purpose was to encourage users to take exercise, interact with them on a daily basis, remember previous interactions, participate in small talk and respond with apparent empathy to user activity.

A highly imaginative, virtual world example is the ‘Gay Cruise’ application [19] which encourages the use of condoms among promiscuous homosexual men. Users choose a ‘purser’ (their relational agent) who guides them through various scenarios on a virtual cruise ship instructing the user on HIV prevention. This is an automation of a health promotion technique known as Intervention Mapping [29,30]. The details of Intervention Mapping are not important to this review, but it is significant that due to the purser being a relational agent the whole process, with relationship building as a critical component, can be orchestrated in a virtual world by a computerised agent instead of a human being. A non-relational computer agent would not have been able to do this. This concept is worthy of further consideration; if agents are created that can build long term relationships with users, procedures normally orchestrated by people (in which a certain level of interpersonal relationship is considered critical) will become eligible for automation. Similar plans are underway to create a ‘stop smoking’ coach [27], once again, a relational agent orchestrating Intervention Mapping. Other examples of behavioural change procedures orchestrated by relational agents are Motivational Interviewing [31,32] and theTranstheoretical Model [33,32].

In other-health based examples, relational agents are employed to serve as counsellors who comfort and empathise with people undergoing medical interventions [32], help explain complicated medical documents to people who are not ‘health literate’ [15], teach social skills to children [34], and proactively interrupt a subject’s day by reminding them to rest, to take medication or to exercise [20-22].

The example of a relational agent with arguably the highest ‘cuteness’ factor is PaPeRo [25]. This is a pet robot that remembers up to 10 faces and displays excitement when it recognises someone. It tells the weather forecast through a Wi-Fi internet link, dances when music is played and blushes when kissed. It also obeys a range of varied commands and displays emotions such as embarrassment, affection and delight. It even attempts to proactively adapt it own personality to be like its owner, introducing such traits as worrying about its appearance. PaPeRo is one example of a ‘sociable robot’ that remembers people, other examples are NeCoRo [26], a robot cat that likewise seeks to build a relationship with its owner and Zeno who is an animated character ‘brought to life’ by Hanson Robotics [35]. To better conceptualise sociable robots, the reader is guided to enter some of these names into YouTube for some quite remarkable footage. In yet another behavioural change example, MIT have now developed a sociable robot [24] that lives with the patient and employs relational agent techniques to provide long term social support for obese people undergoing weight loss programmes [36,37].

The i-cat [23] and major-domo [16] are ‘ambient intelligence’ examples of relational agents. (Ambient intelligence is embedded technology that surrounds users in their home [39]). The i-cat and major-domo are agents that interface to household technology and perform actions like ‘putting on’ some music, turning the TV off, checking for e-mails, advising on product purchases, checking the contents of the fridge, creating shopping lists, programming the washing machine and so forth. In both examples, the agents have many human features. The i-cat is a 38 cm tall robot that looks like a cat’s head; it has robotic facial expressions such as happiness, sadness, disgust and anger to name a few. It recognises users and remembers little details about them like their name, login details, passwords and so forth. Major-domo performs similar actions to the i-cat but its physical form is very different being a virtual butler, a computer character that resides on a series of large interactive screens around the home. From the relational aspect, major-domo goes a stage further than the i-cat by aiming to recognise user habits and proactively offering services.

Tinker, on the other hand, is a virtual museum guide. Found at an information point on a large screen ‘he’ uses a biometric identification system to recognise users previously interacted with [17,18]. Operating as a standard information point, he provides users with information about the museum. However, when users return to the information point he remembers what he previously told them and will comment on this, either on the same visit or upon the user’s return at a later date. His ability to remember users and refer to their previous interaction causes a lot of surprise among subjects as they assume he is a standard non-relational information point. Human responses to relational agents are discussed in a later section but it is important to note that Tinker is not introduced to users as a relational agent but as an information point hence the surprise when users realise that Tinker remembers them.

Modern computer and video games, particularly role playing games (RPGs), contain a lot of high quality agents. The constant improvement of such agents is rapid and it is now quite normal for players to develop a form of relationship with such characters. The manner of the relationships formed and the techniques used are as varied as the games themselves. Observing this speed of development, it is apparent that the routine employment of relational agents as non-player characters (NPCs) in computer games, especially online RPGs, cannot be far away.

The focus (not surprisingly) for NPC development however, has been recording information about interactions as it relates to the game play. In a fighting game for example a tendency to use a lot upper body shots might be recorded to make a repeat fight with the same character more realistic; this type of memory can be seen in NHL 2009 [40] where past experience enables the opponent NPCs to anticipate a player’s next move. It is also common in massively multiplayer online role-playing games (MMPORG), such as World of Warcraft, for users and NPCs to affiliate to a ‘faction’, a “game-defined collection of NPC characters who share an affiliation with a race, government, or organization. By completing quests and killing NPCs, players can affect their standing with a particular faction, causing members of that faction to treat the player differently” [41]. In all such cases a range of values are stored about a user or NPC which can be affected by the user’s actions and then referenced by NPCs. The range of values stored can be complex as can the hosting data structures [for example 42] which, in some cases, employ artificial intelligence (AI). The norm in computer games, however, is for user actions to affect a set of values that will later be referenced by NPCs. Specific activities and conversations typically are not remembered in their own right; there is no express purpose of relationship-building over the long term.

The many techniques used in computer games demonstrate the vast spectrum of relationship-related memory techniques that agents in any environment can use. This spectrum ranges from a single number that dictates an agent’s ‘affection’ for a user, to a full database or AI engine that exists for the purpose of recalling previous interactions and evolving a relationship. At exactly which point an agent becomes a ‘relational agent’ is not precisely defined and as agents everywhere evolve towards the existing relational agent definitions the lack of clarity around this boundary will be more obvious. A current example of this is AlphaWolf [34]; this educational game proposes to teach social skills to children. Focused on cubs finding their way within a pack of wolves, it presents an advanced computational model of social relationship formation. What is not clear from the publication is the extent to which this progression is over successive days or logins.

As it stands today the majority of computer game examples cannot be described as relational agents. It is a reasonable observation, however, that findings related to social agents in any HCI context are analogous to those operating in computer games. This underlines one of this articles key propositions, that relational agents are starting to emerge in a range of virtual worlds and many existing agents are likely to be developed as relational agents.

The relational agents used in behavioural change applications were designed to be persuasive and it was hoped that they would persuade users to adopt healthier lifestyles. From the outset, this became questionable. In the first example [9,12], Laura attempted to encourage exercise but when operating with 101 young people, 69% of whom were MIT students, she was ineffective when compared to the control; the relational agent’s subjects did not do more exercise. Despite this failing, significant positive findings came from this research in that a substantial majority of subjects ‘liked’ the agent and wanted to carry on working with her as compared to the non-relational control. Although some people did not share this experience, unable to conceptualise the idea that ‘she’ was anything more than an algorithm, the majority of participants felt that they did have a relationship with Laura and they even felt cared for [11]. These results were echoed in a similar study [10] that also operated with young people (91 participants with a mean age of 25). Although the majority of subjects warmed to the character there were a few who did not, providing such responses as ‘Laura is not a real person, and therefore I have no relation whatsoever with her’ [10]. Overall, this suggests that for the majority of people, the relational qualities of the agent made the application significantly more attractive and enabled some kind of bond (a relationship) to be created between the user and the character.

In these early experiments, many subjects reported an annoyance with the relational aspects of the character (small talk, facial expressions, for example) after the initial novelty period wore off, primarily due to its repetitive nature. It is speculated [13] that this was more of a limitation of the Laura character than the relational agent concept. Encountering a real person who insisted on similar small talk at each interaction would also be highly tedious and aggravating. It was further speculated that this annoyance would deepen over time but in a rather unusual twist this proved to be false; in a later and longer experiment [13,14], the relationship actually deepened over time. In this later experiment, working with older adults (average age 74), Laura won the affection of her users quite remarkably. On a scale ranging from stranger (1) to close friend (7) the mean came to 6.8. These results are reflected in the comments that the users made about the system being nearly all positive and demonstrating elements of a human-like relationship with the character. These experiments also proved to be persuasive with the relational agent subjects doing much more exercise than the control group. Although participation with the system decreased from an average of 4.6 to 2.3 contacts a week at the end of the 60 day trial, participants indicated they would like to continue using the system with an average weighting of 6.4 (1=”not at all” and 7=“very much”). These results are significant and strongly support the arguments that relationships can be built and that relational agents make applications more attractive to use. The other initial experiments [9,12] also supported the relational agentargument but there is no question as to the greater success of this later, longer experiment.

The key difference could be the age and social mix of the participants. In the more effective application, the adults were older (average age 74), nine out of the ten scored a low reading literacy, all were female, half had never used a computer before and they were nearly all (90%) African Americans. A considerably different group to the young educated people who participated in earlier trials. Any of these differences could be responsible. Consider for example the ethnic mix – one participant reported they liked the fact Laura was a “person of color” [13]. Various studies have suggested that automatic social response and stereotyping are as applicable to human computer relationships as normal communication between humans [44,45] and that are perceived to be more attractive communications from agents of the same ethnicity as the user are perceived to be more attractive, trustworthy, persuasive and intelligent [46] – all important components of human relationships. Interestingly, a number of applications now enable the user to choose their own relational agent [19,47] and include the option to select an agent they are naturally attracted to in some way.

Personality types were found to modify the response that users had to REA the multimodal virtual estate agent [8,10]. Extroverts and introverts reacted differently as did passive and active subjects. Extroverts for example trusted REA more in embodied interactions but, over the telephone, being an extrovert had no impact. In face-to-face contact, introverts trusted the purely task-orientated version more while extroverts trusted the agent significantly less in this mode. Passive subjects felt more comfortable interacting with REA than active subjects did (regardless of the communication mode) while preferring face to face communication to communication over a telephone. It was hoped that such personality experiments would help to diagnose the factors that dictate if an agent will be trusted and/or a relationship built but more recent work indicates that such responses are not standard with all relational agents reflecting instead the perceived personality of the agent [48]. REA, for example, acts like an introvert and if she were real this would invoke various reactions from various personality types. These findings propose one of two conclusions; either natural inter-personality reactions transfer to the world of relational agents or different ersonality types react to agents differently. Although there could be elements of truth in both conclusions, analysis in this case suggests the former.

When a relational agent was developed for an electronic learning environment [43] designed for children between the ages of 12 and 15 years, the results were mixed with a marginal majority preferring the relational agent to the control. When summarising responses, Gulz is clear that many liked the relational agent while others viewed the small talk, for example, as annoying and unnecessary [43]. This somewhat contrasts with the findings related to Laura and REA [9-14] as Gulz points out. It has to be concluded that the success of relational agents in establishing relationships with people is varied and insufficient work has been done to diagnose the critical success factors. It has already been suggested that this could relate to the socio-Human Response and Agent Capabilityeconomic groups involved, the similarity between the agent and the user, personality types or the IT literacy of the user. Looking at this educational example [43], it might also be suggested that age is a factor. Although these experiments employed different agents and quite different scenarios, a trend can be observed. Gulz, when working with children found a 50/50 split in the establishment of a relationship with the character. Operating with young adults, Bickmore [9,10,12] found that the majority were able to establish a relationship with the agent, but when working with older adults, 100% of participants established a relationship [13].

It has already been observed that relational agents are capable of performing tasks designed for human orchestration [19,27,32] and that when it comes to persuasion [9,12] their effectiveness as compared to the control is dubious. In another example, where an agent was used to explain complicated medical documents to the ‘non-health literate’ [15], the effectiveness was once again questioned and the relational agent’s subjects did not understand the literature any better than when they interacted with the control, which in this case was a real person following a similar script to the agent. In these examples it should be clarified that on no occasion were the relational agents found to be inferior to a control. However it would be a mistake, without further evidence, to conclude that relational agents are therefore more effective at performing tasks than people or non-relational agents.

As previously discussed, despite not improving the outcome of a procedure, relational agents do make a number of new activities eligible for automation. It is also clear that relational agents attract user participation. In the case of agents being employed to explain complicated medical documents [15], the subjects actually preferred the agent to dealing with a real clinician (which, as mentioned above, does not mean that the agent was necessarily more effective in the task of conveying information than the clinician). One apparent reason for this is a perception that clinicians are intelligent and busy people and that a patient has limited time with them and does not want to display ignorance and the consequent loss of face. With the relational agent, time is not limited and ‘face’ is not at risk. Another example supports this argument, when a relational agent was employed as a museum guide [17,18] a significantly smaller majority of respondents said they would prefer to interact with the agent than a real person, perhaps because the public perception of museum guides is typically less deferential than clinicians. It is reasonable to surmise therefore that there is a correlation between how important a person is and how much people will welcome a relational agent imitation of them performing aspects of their role. It is also significant that in the case of the museum guide, although the majority of people who preferred a relational agent over a human was smaller than the clinician case, this was still a majority (56% preferred the agent to a real person, 31% said they would prefer a person and 13% were unsure). It appears that the most people simply ‘like’ interacting with relational agents, certainly more than interacting with other agents and in some cases they are preferred to real people.