Social Attitudes and Personalities in Agents

We describe a system (GOLEM) which is aimed at formalising, implementing and experimenting different kinds and levels of social cooperation, represented by different social attitudes or personality traits. First, we examine why personalities in agents are needed and what they are. Second, we propose our definition of the two basic elements of Multi Agent cooperative activity (Delegation and Adoption), explaining how they are related to the agent’s level of autonomy and of cooperativeness. Then, we present how we formalise these levels of Delegation and Adoption in terms of agents’ personality traits or attitudes, and we outline how they can be organised in reasonable personalities and interesting interactive situations. Finally, we show how these traits and attitudes are involved in deciding what to do both proactively and in response to the other’s social action, and in reasoning about the other’s mind. Why Do Agents Need Personalities? Agents endowed with personalities or personality traits, characters, imtividual attitudes, etc. are spreading around in various domains (Cohen and Levesque,1990; Lomborg, 1994; Hayes-Roth,1995; de Rosis et a1,1996; Cesta et al, 1996; Loyall and Bates,1997). Which are the reasons of this trend.’? Just curiosity, or is this a necessary development of the "agentification" of AI? There axe, in our view, several independent reasons for introducing personalities in Agents. Let’s summarise them. a. Social/Cognitive Modelling One of the major objectives of AI (and ALife) as science modelling natural intelligence. Since in nature and in society agents have personalities and this seems an important construct in psychology, one might aim at modelling personality in agents (or emotions or cognitive biases) to reproduce relevant features of human interaction. h. Believability and entertainment Believability has been recognised as one of the most important features for a natural user interaction in entertainment and user-friendly interfaces. It is strongly related to expressing emotions and caricatures (Loyall and Bates, 1997; Walker, 1997) and to reacting in a "typical", or "peculiar" way. Personalities were in fact first introduced in AI to make more "believable" and deceptive some systems like the paranoid PARRY. c. Story and situation understanding In making the required inferences for understanding a story or a situation, it is necessary not only to know the appropriate scripts and frames and the agents’ intentions and beliefs, but also their personalities. The first quite complete and formal theory of personality was intxoduced by Carbonell (Carbonell, 1980) for this purpose. Carbonell says: "Whenever a story includes character development of one of the actors, this development turns out to be useful and often crucial in formulating an understanding of the story." (p.217). This claim is also connected to what is now called "agent modelling": not only in stories but also in real interactions (in human or virtual reality) "knowledge about personality traits is necessary to understand the actions" of the agents. d. Agent Modelling User stereotypes and profiles proved to be useful in adaptive and cooperative human-machine interaction, to make correct ascriptions and abductions (Rich,1989). The same is true multiple agents’ interaction. We therefore need defining agents’ classes and stereotypes, some of which are personality-based. For example: in user modelling, student denotes a role, whereas aristocratic or thrifty denote personality traits. Among agents, we might have classes like mediator or executive agent or information filtering agent, but also classes like benevolent or self interested, which correspond to social personality traits or attitudes. All these are interesting reasons for introducing/modelling personalities in the agents. But we believe that there is some more principled reason that holds in the very basic philosophy of agent-based computing: its decentralised character, its open world assumption (Hewitt, 199&), its "experimental" approack e. Exploring and comparing strategies One of the most interesting aspects of decentralised and MA systems is that the3’ provide a scenario for experimental exploration of coordination mechanisms, behavioural strategies and organisational structures which could not be designed or predicted by centralized rational planning. This is crucial in the "open system" perspective that characterises the new AI of the ’90s (Bobrow, 1991; Hewitt,1991). Exploring different behavioral, reactive or planning strategies in multiagent systems can be seen as exploring adaptivity, efficiem’y and coexistence of different "personalities" in agents’ models. Personalities had already been introduced in the different kinds of commitment defined by Cohen and Levesque (Cohen and Levesque, 1990) orby Rao and Georgeff (Rao and Georgeff, 1991). No strategy can be defined as a-priori optimal, since the world is open, it changes, it is uncertain and unknown, and since other agents in the world will adopt strategies that 16 From: AAAI Technical Report FS-97-02. Compilation copyright © 1997, AAAI (www.aaai.org). All rights reserved.