Reply to Comments by Bainbridge,Gratch,and Nishida

Both Bainbridge (2012) and Nishida (2012) raise issues in relation to embodiment. Bainbridge references the work in mirror neurons, considered a possible mechanism through which theory of mind can function (Rizzolatti & Craighero, 2004). Nishida asks whether one should add modelling at the physiological level, in order to take account of the somatic marker theory of Damasio (1994), rooting affective responses in brain–body interaction.

In our work (Aylett & Paiva, 2012), affective responses are generated through the interaction between the perception of events by an agent and its current internal state, modelled as a set of goals, drives, affective variables, and plans. While this incorporates feedback between modelled cognitive processes and affective state, it does not currently include interaction between embodiment and affective state. There is substantial work in the modelling of neuro-physiological processes relating to emotion covering a multiplicity of brain areas, as well as important steps towards modelling the role of a mirror neuron system in agent action-selection (Thill & Ziemke, 2010). However, these models are so far very specific and low-level, using nonsymbolic representations such as artificial neural nets. They usually relate to a single emotion or type of action, making it particularly hard to see how they could be interfaced to a symbolic-level model such as ours. Without a generalised account of how symbolic representations can be generated from these lower level processes—way beyond the state of the art—it is not clear how these models in their current form could be integrated into an agent that interacts at the level of natural language.

That said, relevant work has taken place in robotics research where embodiment issues are unavoidable. In a more recent project, we have taken the model we discuss into a robotic context and are currently investigating what mechanisms can be used for feedback into the FAtiMA-PSI model from nonsymbolic processes (Castellano et al., 2008).

The second theme emerging from the comments is articulated in some depth by Gratch (2012). This is the extent to which behavioural properties should be modelled explicitly and declaratively rather than generated dynamically by interaction between the agent and its environment. In artificial intelligence, the idea that complex behaviour can be an emergent property of such interaction is especially associated with the mid-1980s work of Brooks (1991). Internal complexity is removed and the agent becomes both more robust and more responsive, while its behaviour becomes concretely situated rather than abstractly specified. This relates strongly to embodiment, as it is through the body that much of this interaction takes place.

Similar discussion has taken place outside of agent computational models, and notably in social science. As Bainbridge points out, intense debate surrounds terms such as culture and personality, and the same may be said of emotion. The recent growth of neuropsychology and criticism of the previously influential cognitivist approach underlying some of the theory we have implemented in our model would also include an interactional account of behavioural complexity as one of its elements. Gratch correctly suggests that linguistic terms used descriptively in psychology can be misleading if slavishly implemented in a generative system such as an agent architecture.

We also subscribe to the importance of this idea and our model deals with personality via interaction. Rather than representing personality explicitly, it emerges as patterns of behaviour over time derived from the threshold and decay parameters set for the emotions modelled and their impact on the actions selected by the agent. For this reason, incorporating the Big Five personality dimensions, as Bainbridge suggests, would not be easy. The Big Five model offers a descriptive set of linguistic terms representing stable features of behaviour over time, which necessarily take place in a variety of contexts. Our model would generate the behaviour with reference to these contexts and the current setting of the emotional parameters in relation to the current goals and action repertoire of the agent, rather than through a rule about what an agent high in neuroticism might choose to do.

We take Gratch’s point that one could apply a similarly emergent approach to culturally specific behaviour. He argues that culturally specific behaviour depends on the environment and not just on the choices of an agent, and that in the human case, there is evidence that social pressure is at least as strong a mechanism for producing this behaviour as individual commitment to cultural norms.

We do not dissent from this idea in principle. In the model we presented, agents can be seen as having successfully internalised the cultural norms and as always using these to determine their contextual behaviour. However, because the model incorporates cognitive appraisal, changes in context—or the impact of the environment—are necessarily included. In mapping the descriptions of the Hofstede individualism–collectivism dimensions onto the parameters of the model, social behaviour—in interaction with other agents—will depend not only on the cultural parameterization, but on whether the agent likes or does not like the agents it is interacting with. Agent goals are not only associated with praiseworthiness (the degree of social value), but also with desirability (the degree of individual value). Thus it is perfectly possible under certain circumstances for an agent to abandon a socially valued goal in order to pursue an individually valued one.

Nevertheless, we agree that the model does not include any specific mechanism representing the degree of social pressure to conform, though this is potentially implementable and would be an interesting direction to pursue. We would accept that the model we present is not yet sufficiently social in its scope: starting from the requirement for an agent-centred mechanism for selecting actions, it displays an excessive degree of autonomy in its current form.

Finally, Bainbridge raises some of the ethical issues of working in this area in commenting that cognitive empathy could be a mechanism for manipulating others as well as empathizing with them, and that a technology for teaching tolerance could be used to teach the opposite. The domains to which we have applied our model have made us very conscious of these issues, though aware too that we share them with other research fields. The only defence against immoral uses of technology lies in the social mechanisms of the society in which it is developed, and this must start with the technology developers themselves.