Bazzan and Bordini’s research into the integration of emotions in computer science is presented in [12] where they discuss the emotional model they have implemented, the details of the simulation constructed to investigate their research questions and the results of these simulations. Bazzan and Bordini use the OCC model (discussed in section 4.1.1.1) exclusively to model emotion. Their justification for this is that the model groups emotions according to their eliciting conditions which facilitates computational implementation since there are some pre-existing distinctions. Bazzan and Bordini model four emotions in [12], the names of these emotions, their eliciting conditions and effects upon intentional behaviour of agents are listed below:
Joy: elicited if an agent has collected a certain number of points or if at least a certain number of neighbours are joyful. Elicitation of joy in an agent causes the agent to cooperate.
Distress: elicited if an agent has not collected a certain number of points or if at least a certain number of neighbours are distressed. Elicitation of distress in an agent causes the agent to defect.
Pity: elicited if a neighbour has not collected a certain number of points. Elicita- tion of pity in an agent causes the agent to cooperate.
Anger: elicited if an agent has not collected a certain number of points and the opponent has defected. Elicitation of anger in an agent causes the agent to defect.
When an emotion is elicited in an agent, the target of this emotion is the agent’s intentional behaviour which is altered so that an agent may increase its monetary utility; emotions are therefore functional. In [12], emotions are implemented as distinct values and their effects are manifest when this value (potential) has reached or surpassed a user-set threshold. These emotion potentials are intended to be calculated in much the same way as prescribed by the OCC model i.e. local and global intensity variables are taken into account and such variables are intended to be domain-dependent. Emotion modelling is therefore intended to be quantitative.
In the simulations used in [12], Bazzan and Bordini only endow agents with one emotion in the same vein as Oliveira (see section 4.3.6). Consequently, opposing emo- tions may not be experienced in context of the same emotion-eliciting event. Other non-emotional agents in the simulation may cooperate or defect depending upon the strategy they have been programmed with (details of these strategies are not provided in [12]). Furthermore, the percentage of initial cooperators in the population can be set by the operator. The simulations themselves take place on a two-dimensional, thirty by
thirty square grid with each agent occupying one square each (ninety agents exist in the environment at any one time). The boundaries of the environment are fixed and each agent plays against its eight immediate neighbours (four to the north, south, east and west of the agent and four to the north-west, north-east, south-east and south-west).
In each round, every agent will play a Prisoner’s Dilemma game against its eight op- ponents with the highest scoring agent from the neighbourhood taking over the squares occupied by agents that scored lower. This style of play is intended to build upon work performed by Nowak and May [135] who also simulate an iterated Prisoner’s Dilemma game where the strategies of successful players are propagated into neighbouring cells. The findings of Nowak and May’s research illustrates that different equilibria will arise in different neighbourhoods as any one player does not play against the entire population. Therefore, Bazzan and Bordini wish to investigate if emotion may be able to disrupt this equilibria in neighbourhoods so as to promote cooperation in neighbourhoods of defec- tors. The research may then be described as being driven by the cognitive-engineering motive according to Burghout et al.’s classifications [25].
As Bazzan and Bordini note themselves in [12], their framework is intended to be generic so that it may be used as a starting point by those wanting to generalise the rules associated with the emotions implemented. Therefore, the functions responsible for calculating potentials of emotions are explained quite generally i.e. whilst the variables to be used in such calculations are listed, the calculations themselves are not. Another feature of the framework is that agents are guaranteed to perform the behaviour pre- scribed by an emotion if that emotion is elicited. This design choice echoes all other computational models of emotion discussed in this section with the exception of Reilly’s emotional model (see section 4.3.1).
4.3.8 “SHAME+”
The SHAME and SHAME+ systems that are presented respectively in [149] and [25] by Poel et al. and Burghouts et al. are examples of a MAS whose aim is to test the effectiveness of the emotional model developed. I will focus upon [25] since this paper provides details of the augmented version of the SHAME system. That is not to say that I will ignore [149] entirely, instead, if any references are made to this paper they will be indicated clearly to avoid any potential confusion between the two works.
Based upon the research goal outlined above, it can be asserted that the research motivation of Burghouts et al. is a part of the believable-agent motive since they wish to conduct research into how best to model emotion so that natural human interaction is possible (the authors explicitly place their work into this research motivation category themselves). Furthermore, the resulting social interactions between agents are investi- gated to consider the effects emotion may have upon this facet of an agent’s operation. To achieve this goal, a model that relates emotional and cognitive processes with be- haviour is required so that an agent can be perceived as performing naturally by way of expressing emotions and particular personality traits. To study the effect emotions
have on social interaction, some form of test-bed is required which incorporates such a feature; in [149] and [25] this is achieved by constructing a bespoke simulation.
Fourteen emotions are included in the SHAME and SHAME+ systems (since there are so many they will not be listed here) and are modelled as opposing pairs. Each pair shares one scalar value that represents the potential/intensity of the opposing emotions. By doing this, Burghouts et al. ensure that joy and distress, for example: cannot be active at the same time with respect to the same event as they are antipodes of the same scale. A positive (negative) intensity value on an emotion’s scale corresponds to the positive (negative) emotion represented.
Emotion elicitation is triggered by an agent appraising an event in the simulation based upon methods described in the OCC model of emotion4 (see section 4.1.1.1) but are not exact replicas. As inferred by the discussion of how emotions are implemented by Burghouts et al., emotion elicitation is quantitative and is driven by the occurrence of events that are external to the agent rather than events that are internal (homoeo- static events, for example). Simulation events are mapped to emotional states for each type of agent: therefore events are appraised by agents but this appraisal stops after identification of the type of event. Ergo, appraisal is not implemented exactly in the way envisioned by the OCC model [142] where an emotion is elicited by the agent assessing how an event may impact upon its goals, standards and attitudes. It would appear then that events cause a pre-determined positive/negative alteration of the emotion scalar it is intended to affect. So, an event such as “attacked by predator” may subtract ten from an agent’s joy-distress scalar, for example.
For all emotions implemented in SHAME and SHAME+, activation thresholds are set to zero by default. Therefore, when an event alters the scalar value for an emotion pair, the positive/negative emotion represented by that scalar value is activated. The intensity of the emotion experienced is determined by calculating how far from zero the current emotion’s potential has shifted e.g. if an agent’s joy/distress scalar is altered positively by ten then the agent will experience joy with intensity of ten. There is an issue with respect to this method of implementation as it is extremely unlikely that an agent may be neutral with respect to an emotion pair (their shared emotional scalar must equal exactly zero for this to be true). An important feature of the emotional model implemented in SHAME and SHAME+ is that all emotional scalar values are bound at -100 and +100 so the intensity of any emotion may not continue to increase indefinitely. Essentially, such a feature implements the concept of a saturation point (first considered and implemented in Cathexis, see section 4.3.3) and is important to note as it reflects a realistic, human-centric emotional model.
The SHAME and SHAME+ systems are most notable with respect to this thesis by virtue of their consideration of emotional personalities. In context of SHAME+, personality defines emotional characteristics i.e. what emotions an agent can experience, how strongly/weakly they are experienced when elicited and how long the emotion lasts
4
Details of the appraisal process are taken from [149] where they are described in more detail than in [25].
for when elicited. The definitions for “strong” and “weak” are ambiguous in this case; it may mean that the threshold for an emotion’s activation is lower or it may mean that a certain event type causes a greater shift in emotion intensity. To the best of my knowledge, these definitions are not clarified in either [149] or [25].
Turning to the issue of emotional effect it should first be highlighted that an emotion affects an agent’s intentional behaviour and the intensity of an agent’s emotions does not affect the behaviour exhibited. The question may then be asked, what is the pur- pose in implementing emotional intensities if they result in non-variance of intentional behaviour? This question may be answered by considering the decay function utilised by the SHAME and SHAME+ systems which makes use of knowledge provided by the user to alter an emotion’s intensity over a period of time. Unfortunately, unlike Stuenebrink et al.’s, Reilly’s and Vel´asquez’s emotional decay functions (see sections 4.2.1, 4.3.1 and 4.3.3 respectively), details of how the function operates are not discussed. It is therefore difficult to assess whether this function could be adopted/modified for my own purposes or if it has any sound psychological foundation.
Burghouts et al. also model “self-control” which removes or reduces the dissonance between expected behaviour and behavioural standards. The mechanism used for this step is inspired by Bandura’s observation [8] that emotional self-control and the potential actions that may result from particular emotions being elicited includes self-monitoring via. personal standards and corrective self reactions. Therefore, the emotion’s effect cannot be said to be entirely guaranteed as this self-control may suppress the associated action. Justification for including a mechanism for self-control is provided by Gifford [96] who states that self-control is required to reduce conflicts between present and future gratification. The calculations and functions for applying self-control are not outlined in [25] therefore, it is difficult to ascertain how this should work and consequently, I am not able to adopt/modify such a function for this thesis.
4.3.9 “EPBDI”
Zoumpoulaki et al. introduce theEmotion, Personality, Beliefs, Desires and Intentions model (EPBDI) in [216]. Note that this model is not a direct extension to the EBDI model proposed by Jiang et al. in [95] (see section 4.3.5). The important difference between EBDI and EPBDI is the concept of personality which was first mentioned in section 4.3.8. Zoumpoulaki et al. computationally model personalities using theOCEAN model (discussed below) whilst emotions are modelled using the principles set out by the OCC model [142]. In [216] it is proposed that, by modelling personality and emotion, the required mechanisms for simulating realistic human-like behaviour under evacuation can be outlined and verified.
To this end, Zoumpoulaki et al. implement the following emotion pairs using the OCC model as a basis: joy/distress, hope/fear, pride/shame, admiration/reproach and sorryFor/happyFor (first three pairs affect the agent that experiences the emotions, the last two affect other agents). Given the discussion in [216], these emotions appear to be
modelled in much the same way as emotions are modelled in the SHAME and SHAME+ systems (see section 4.3.8) but is neither confirmed or denied by Zoumpoulaki et al. Since emotion pairs are modelled using scalar values (as prescribed by the OCC model) it would appear that emotions are modelled quantitatively in [216].
The emotions modelled act functionally to affect the perception of an agent, the agent’s appraisal of events, an agent’s decision-making and the actions performed by an agent in a given situation to cope with that situation in a particular way. The consequences of specific emotions upon these components along with whether or not the effects of emotions are guaranteed is unfortunately not made clear. Despite this, it is important to remember that emotion has some effect upon the agent’s decision-making and the intentional actions subsequently performed. Since emotions are modelled as opposing pairs and appear to be driven by events/appraisals, it would be reasonable to assume that opposing emotions may not be active with respect to the same eliciting condition although confirmation of this is not made explicit in [216].
An agent’s personality affects how the agent appraises events and its decision-making that results from this. As stated earlier, an agent’s personality is modelled using OCEAN, an acronym of characteristics that stands for: openness, conscientiousness, extraversion, agreeableness and neuroticism. However, like the implementation of emo- tions discussed previously, insufficient implementation details are given in [216] with regards to how personality influences appraisal. Therefore, these ideas have not been implemented in the model used in this thesis.
Zoumpoulaki et al. make use of a simulation to investigate whether computational modelling of emotion produces more “human” or believable behaviour. Therefore, Zoumpoulaki et al.’s research motivation may be classed as being a member of the believable-agent research motive [25]. In this simulation, agents are situated in a two- dimensional environment consisting of static objects and randomly generated pockets of fire that may hurt or kill agents. The initial agent population’s demographic/personality distribution and the position/spread parameters of fires are all user-defined.
4.4
Chapter Summary
In this chapter I outlined the various ways in which emotions could be modelled by first considering a number of psychological emotion models in section 4.1. Following this I then discussed prominent logical formalisms of some psychological models of emotion (see section 4.2) before presenting a review of literature concerned with how psychological models of emotion have been modelled computationally and used to answer various research questions (see section 4.3).
The literature reviewed in this chapter has led to a number of important conclusions that will serve to inform and direct my own emotional modelling framework. In chapter 5 I will use the conclusions drawn from the discussions presented in sections 4.1, 4.2 and 4.3 above to construct my own research agenda and how I intend to answer the research
questions posed. Specifically, this will entail a description of the emotional model I have developed and a discussion of the test-bed that I have used to enable an investigation into the research agenda proposed. Table 4.1 summarises the key points of the research discussed in section 4.3 that are of interest to this thesis. The intention of this table is to facilitate a comparison between key features of this research and my own emotion modelling framework that is outlined in chapter 5.
With respect to these conclusions a number of salient observations have been made. The first is that appraisal models of emotion (see section 4.1.1) are particularly well suited as a basis upon which computational models of emotion can be developed. Of the three types of emotional models considered (appraisal, dimensional and anatomi- cal), appraisal models are the only type of emotional model to both consider the origins of emotion and consider them from a non-physiological perspective. Establishing how emotions are elicited is particularly important since without such knowledge, emotions themselves may never be modelled. The non-physiological standpoint is also exception- ally beneficial as computers are obviously incapable of physiological responses. Partic- ular attention should be afforded to the OCC model of emotion (see section 4.1.1.1) since this appraisal model was developed with computational implementations in mind. This makes the OCC model especially tractable and attractive from a computational standpoint.
Aside from these observations, the review of literature in this chapter has highlighted an important nich´e in which my own work can exploit so as to produce novel, interesting research. The majority of research presented in section 4.3 is motivated by cognitive- engineering research proposed by Burghouts et al. in [25]. This field of research is quite interesting since there are many avenues of enquiry still unexplored. As Bazzan and Bordini state in [12]:
“..little work has focused on the investigation of interactions among social agents whose actions are somehow influenced by their current emotional set- ting.”
My own research will therefore be conducted in the context of this motivation as there are good foundations for the design and implementation of such research whilst there are still ample opportunities to produce novel work.
A succinct summary of the salient observations considered in this chapter can be found below.
The most researched and appropriate general psychological emotional model for use in computer science appears to be the appraisal model of emotion.
Logical formalisms have focused upon appraisal models, most notably the OCC model. Five of the nine implemented computational emotion models discussed are also based upon the OCC model.
T able 4.1: Key p oin ts of researc h discussed in section 4.3. Researc h Public Go o ds T estb ed? Psyc h. Mo del Used Emo. Elicit Quan./Qual.? Emo. P ot. & Act. Thresh.? Multi. Emo. p er Agen t? Emo. Mo delled Emo. Affect Emo. Ch.’s? Emo. De- ca y? [155] Sec 4.3.1 No [142] Quan. Y es Y es OCC’s 22 + resen tmen t, frustration, startle Unin ten tional & in ten tional b eha vi our No Y es [68] 4.3.2 No [67] Quan. & Qual. Y es Y es Unkno wn In ten tional b eha viour No No [205] 4.3.3 No Man