Behind the Scenes of Life-like Personal Service Assistants:
A Structured Approach to Integrating Personality Plans in Agent Behaviours

Yasmine Arafa, Patricia Charlton, Abe Mamdani

Intelligent Communication Systems, Imperial College of Science, Technology & Medicine, UK
Exhibition Road, London SW7 2BT

Synthesising the metaphor of life-like visual Personal Service Assistants (PSAs) is the focus of much current research. Our research goes behind the scenes of stage representations and approaches the technical infrastructure enabling the creation of more believable life-like interface characters. The major challenges in building such agents are not only getting them to intuitively interact with the user and the surrounding environment, but also, planning and visualising their intuitive behaviour through context sensitive character traits.

Animated life-like character representations of PSAs will only be widely deployed when there are real-time mechanisms to contextually map the personality models to affected personality representations. To approach this we propose the following design elements of an architecture for including personality and emotional responsiveness in a PSA: a) semantic abstraction of the knowledge maintained and exchanged; b) consequently, mapping the resulting semantic understanding into appropriate initial planned behaviour; and c) translating this behaviour into context-sensitive character traits by varying the planned response using variable emotion indicators, then finally representing this in a selected modality(ies). We are investigating the behaviour on both the internal and external (visual or other modality) levels, as individual agents and as part of a community of agents. Aiming to evaluate the effect personality-based behaviour on interactive user interfaces as well as on other service agents in a multi-agent environment.

The functionality of a PSA is not standalone, it is rather that of an agent communicating and collaborating within a community of service agents in a multi-agent system (MAS). Development of agent systems appears to place emphasis on either service delivery or human interaction. It is inevitable that the MAS will have to deal with both, so the effort to arrive at a solution to provide support on both levels means that a structured way of integrating the two demands to be approached. For the KIMSAC system a structured integration of MAS and multimedia was essential to meet the needs of a PSA interface and public information and service access. An MAS architecture was used to supply services while multimedia was used as the main technology to provide the user interface. The structure was based on the design of a language called ADL (Asset Description Language).

For life-like interface agents, like the Personal Service Assistant, to be effective they must exhibit some form of intelligence in the sense that they are intuitive, proactive and responsive to user preferences and requirements as well as to the world as a whole. To be capable of this the PSA must have a semantic and contextual understanding of the information being exchanged. This requires a theoretical framework for representing knowledge and belief of agents interacting with other agents. This includes frameworks for representing uncertain knowledge about the surrounding environment evolving with experience and time, awareness of the implication of time constraints, and context-based behaviour.

For this purpose we define so called Assets which are annotations (meta-level descriptions) of objects being manipulated between the visual framework (PSA) and the backend system (service agents). This provides an understanding of the content being handled and hence a better awareness of the environment as we discuss in further detail in the full paper. These abstractions hold physical and conceptual meaning and also include expected initial emotional indicators. We are examining criteria and ways in which these indicators may vary according to the context of agent conversational acts. With an aim of using this as a base for building agent’s social and adaptive behaviour.

These assets and the Asset Description Language offer:

The template essentially grounds four important aspects of information sharing between agents and the end user. The aspects defined to allow the sharing of content across a distributed system and facilitate scalable service provision are: These models include indicators to emotions expected when a defined asset is being handled in a normal or non-context sensitive way. These indicators are associated, using a set of selection rules, with a predefined intended behaviour. This planned behaviour is then modified to incorporate context-sensitive variables to select the appropriate personality-based behaviour. The resulting behaviour is then mapped to a visual (animated) representation. Variations of the animated (or intonation) can be realised by altering and adjusting the intensity of the behaviour according to the immediate situation.

In our research so far, we assume that a character need not exhibit all the attributes of the personality definition to be a successful personality model. It needs to incorporate at least some of the more basic features of personality. We try to model the broad qualities of personality that include individually distinctive, and consistently enduring yet subject to influence and change. To realise a Personality model, we define a finite set of Personality Types and Manifestation Styles that collectively serve to define the overall representation strategy by which the individualised behaviour a chosen PSA character can be visualised. These manifestations will define the various modalities in which personality may be presented. For simplicity, we concentrate on the visual (motional and gesture) manifestations, but can later extend this to include other modalities like synthesised speech.

The emotion(s) that results from individual situations will influence how the representation strategy will be defined and hence generating personality based on the presupposed personality type. In order to support dynamic creation of representation strategies for visual character-based manifestations there needs to be a semantic understanding of the content being handled in accordance with the context within which it is being conveyed. For this purpose we use the Asset descriptions which include Provisional Indicators to an expected emotion that a given content may convey. We propose, to start with, a limited number of basic (fundamental) emotions that can be anticipated when handling a particular asset. These indicators are negative, positive or neutral.

These indicators alone can not determine behavioural reactions. The impact would depend on past experiences (in general, as well as, experiences with particular individuals), the current mental state and the environment. So the actual influence of such indicators will be determined by these variable factors that are governed by the context in which the content is being conveyed. Nevertheless, these indicators serve as the innate knowledge an agent may start with, and based on which may influence its behaviour. If we are effectively able to use such variables in defining the context we may effectively determine the attitude form and visual realisation of a PSA’s character.

Therefore, a further requirement to the PSA’s perception is not only an understanding of content but also and understanding of the context and how this context may influence the behavioural attitude. Social interaction realised through the conversational acts between agents and the PSA further affect behavioural responses. The communicative intentions can form a basis on which the PSA may infer social and behavioural attitude. This may be structured by associating intentional inference to the constructs when its constraints are satisfied. The PSA must maintain its memory of past dealing with agents and their social behaviour and attitudes towards him. This memory serves to influence or constrain future social interactions with familiar agents, allowing the PSA to alter its own behaviour in a manner in accordance with the other agent. Changes in attitude and resulting social emotion will influence the representation strategies appropriate to interpret this emotional behaviour in accordance with its innate personality.

Online References

[Arafa et al., 99] Y. Arafa, P. Charlton & E. Mamdani, Supporting Personality in Personal Service Assistants, Third International Conference on Autonomous Agents (Agents '99), forthcoming.

[Arafa et al., 98a] Y. Arafa, P. Charlton, E. Mamdani & P. Fehin, Designing and Building PSAs with Personality, 1st International Workshop on Embodies Conversational Characters, 12-15 October ‘98.

[Arafa et al., 98b] Y. Arafa, P. Charlton & E. Mamdani, Modelling Personal Service Assistants with Personality: From Metaphor to Implementation, Workshop on Grounding Emotions in Adaptive Systems, SAB’98.

[McGuigan et al., 98] R. McGuigan, P. Delorme, J. Grimson, P. Charlton, & Y. Arafa, The Reuse of Multimedia Objects by Software in the Kimsac System, In proceedings of Object Oriented Information Systems, OOIS’98.